Skip to content
Dew

Reinforcement learning with verifiable rewards

Open in ColabView on GitHubDownloadGPU

Supervised fine-tuning is easy to explain: show the model the right answer and make that answer more likely. But for a lot of the problems we care about, nobody has written the right answer down. What we do have is a way to check one. Code either passes its tests or it doesn’t, and a maths answer either matches or it doesn’t.

A check is enough to learn from. Let the model try a few times, check each try, and nudge it towards the ones that passed. That’s reinforcement learning with verifiable rewards, RLVR for short, and it’s the recipe behind DeepSeek-R1’s reasoning (DeepSeek-AI, 2025).

The policy writes four replies to a maths question, a checker compares each boxed answer with the reference, and the update makes the replies that scored above the group's mean more likely.questionNatalia sold 48clips in April andhalf as many inMay. How many didshe sell in all?policyP(reply)abcdrepliesreward= 72 ?checkera48/2 = 24, 48+24 =7248/2 = 24, 48+24 =721b48×2 = 96, 48+96 =14448×2 = 96, 48+96 =1440cApril 48, May 24:72April 48, May 24:721dhalf of 48 is24half of 48 is240update: replies that beat the group's mean get more likely
The policy writes four replies to a maths question, a checker compares each boxed answer with the reference, and the update makes the replies that scored above the group's mean more likely.questionNatalia sold 48clips in April andhalf as many inMay. How many didshe sell in all?policyP(reply)abcdrepliesreward= 72 ?checker48/2 = 24, 48+24 =7248/2 = 24, 48+24 =72148×2 = 96, 48+96 =14448×2 = 96, 48+96 =1440April 48, May 24:72April 48, May 24:721half of 48 is24half of 48 is240update: replies above the mean get more likely

That picture is the whole method. We’ll build it with GRPO, the variant DeepSeekMath introduced (Shao et al., 2024) and R1 was trained with, and use it to teach Qwen2.5-0.5B-Instruct grade-school maths with Dew’s GRPOObjective.

Setup

Section titled “Setup”
%pip install -q "dewml[cuda13,interop,streaming] @ git+https://github.com/AshishKumar4/dew@dea386da476a5b32f59b6bfef94993ace34f38c7"
install log
  Installing build dependencies ... done
  Getting requirements to build wheel ... done
  Preparing metadata (pyproject.toml) ... done
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 604.4/604.4 kB 2.7 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 77.7/77.7 MB 22.4 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 532.5/532.5 kB 47.3 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 206.6/206.6 kB 21.9 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 91.4/91.4 kB 10.1 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 3.3/3.3 MB 121.4 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 53.9/53.9 MB 31.8 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 2.2/2.2 MB 100.1 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 216.9/216.9 kB 24.7 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 137.7/137.7 kB 15.0 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 102.6/102.6 kB 11.5 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 8.0/8.0 MB 33.5 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 134.2/134.2 MB 9.5 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 89.9/89.9 MB 22.2 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 56.6/56.6 kB 6.2 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 168.1/168.1 kB 18.3 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 32.1/32.1 MB 38.9 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 135.7/135.7 kB 15.1 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 87.5/87.5 kB 9.9 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 16.0/16.0 MB 53.5 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 109.2/109.2 kB 11.9 MB/s eta 0:00:00
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 115.9/115.9 kB 13.2 MB/s eta 0:00:00
  Building wheel for dewml (pyproject.toml) ... done
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
cudf-cu13 26.6.0 requires pyarrow<24,>=19.0.0, but you have pyarrow 26.0.0 which is incompatible.
MODEL = "Qwen/Qwen2.5-0.5B-Instruct"
PROMPTS = 16          # questions per batch
GROUPS = 8            # replies per question, G
PROMPT_TOKENS = 256   # the longest question we keep, in tokens
NEW_TOKENS = 512      # the longest reply
STEPS = 300           # batches of replies to train on
LEARNING_RATE = 3e-6
BETA = 0.04           # weight of the KL penalty
SEED = 0
import json
import time
from collections import Counter
from dataclasses import replace
from types import SimpleNamespace


import dew  # before JAX's backend opens, so its numerical settings apply
import jax
import jax.numpy as jnp
import numpy as np
from datasets import load_dataset


from dew import LocalTracker, Trainer, figures
from dew.config import OptimConfig
from dew.config.data import PromptData, loader
from dew.data import HFTokenizer, Loading
from dew.data.dataset import DataPartition
from dew.data.prompts import INFO_KEY
from dew.interop import PretrainedDecoder
from dew.nn.backbones.decoder_block import remat_policy
from dew.objectives.base import Step
from dew.objectives.rl import GRPOObjective, PromptRecord, SampledRollout
from dew.rl import clipped_surrogate_terms, k3_kl
from dew.rl.sandbox import MathReward
from dew.sampling import Sampling
from dew.training import prepare_process


np.set_printoptions(precision=3, suppress=True)
prepare_process()  # Dew's setup for a training process, before the first JAX call

A task we can check

Section titled “A task we can check”

GSM8K (Cobbe et al., 2021) is a set of grade-school word problems, each with a worked solution that ends in a single number. We only need that number. We’ll ask the model to show its working and put the final answer in \boxed{}, and Dew’s MathReward gives a reply 1 if its last boxed answer equals the reference and 0 otherwise. It compares numbers exactly, so 72, 72.0 and \frac{144}{2} all count.

What happens when a reply gets the number right but forgets the box?

gsm8k = load_dataset("openai/gsm8k", "main")
train, test = gsm8k["train"], gsm8k["test"]




def reference(row):
    # A worked solution ends with "#### 72"; the number is all we keep.
    return row["answer"].split("####")[-1].strip().replace(",", "")




checker = MathReward()
question = train[0]
print(f"{len(train)} training and {len(test)} test questions\n")
print(question["question"], "\nreference:", reference(question), "\n")
for reply in [r"In May she sold 48 / 2 = 24, so \boxed{72} in all.",
              r"She sold 48 * 2 = 96 in May, so \boxed{144}.",
              r"She sold 48 + 24 = 72 clips."]:
    print(f"{checker(reply, PromptRecord('gsm8k', reference(question))):.0f}  {reply}")
stderr
Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
WARNING:huggingface_hub.utils._http:Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
7473 training and 1319 test questions

Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May? 
reference: 72 

1  In May she sold 48 / 2 = 24, so \boxed{72} in all.
0  She sold 48 * 2 = 96 in May, so \boxed{144}.
0  She sold 48 + 24 = 72 clips.

The unboxed 72 gets the same 0 as the wrong 144, because the checker only reads the answer inside the box. That’s harsh, but it’s what makes the check trustworthy: it never has to guess which number in a reply is the answer.

The model before training

Section titled “The model before training”

We start from Qwen2.5-0.5B-Instruct (Qwen Team, 2024), a small model that already follows instructions and can do some arithmetic. Before we train it, let’s grade it on all 1,319 test questions, which it will never see during training. It answers greedily, always taking its most likely token, so the same question always gets the same reply.

source = PretrainedDecoder.load(MODEL, dtype=jnp.bfloat16, param_dtype=jnp.float32,
                                max_seq_len=PROMPT_TOKENS + NEW_TOKENS)
tokenizer = HFTokenizer(MODEL)
stock = source.text_generation().sampling
ends = dict(eos_token_ids=stock.eos_token_ids, pad_token_id=stock.pad_token_id)
INSTRUCTION = r"Solve the problem step by step, then give the final answer in \boxed{}."




def messages(row):
    return [{"role": "user", "content": f"{row['question']}\n{INSTRUCTION}"}]




def chat(row):
    # The question as the model's chat template writes it, up to where the reply starts.
    return tokenizer.tokenizer.apply_chat_template(messages(row), tokenize=False,
                                                   add_generation_prompt=True)
def grade(variables, rows, batch=128):
    # Every reply to `rows` from the model with these weights, and which ones are right.
    task = replace(source, variables=variables).text_generation(
        sampling=Sampling(temperature=0.0, **ends))
    prompts = [chat(row) for row in rows]
    replies = []
    for start in range(0, len(prompts), batch):
        replies += task(prompts[start:start + batch], NEW_TOKENS, key=SEED).text
    right = [checker(reply, PromptRecord("gsm8k", reference(row)))
             for reply, row in zip(replies, rows)]
    return replies, np.asarray(right)




before, right_before = grade(source.variables, test)
print(f"{right_before.mean():.1%} of the {len(test)} test questions right\n")
print(test[0]["question"], "\nreference:", reference(test[0]), "\n")
print(before[0])
48.7% of the 1319 test questions right

Janet’s ducks lay 16 eggs per day. She eats three for breakfast every morning and bakes muffins for her friends every day with four. She sells the remainder at the farmers' market daily for $2 per fresh duck egg. How much in dollars does she make every day at the farmers' market? 
reference: 18 

To determine how much Janet makes at the farmers' market every day, we need to follow these steps:

1. Calculate the total number of eggs laid by the ducks in a day.
2. Determine how many eggs are sold at the farmers' market.
3. Calculate the total revenue from selling the eggs.

**Step 1: Calculate the total number of eggs laid by the ducks in a day.**

Janet's ducks lay 16 eggs per day. Therefore, the total number of eggs laid in a day is:
\[ 16 \text{ eggs/day} \]

**Step 2: Determine how many eggs are sold at the farmers' market.**

Janet eats 3 eggs for breakfast every morning and bakes 4 muffins for her friends every day. Therefore, the total number of eggs she eats in a day is:
\[ 3 \text{ eggs (breakfast)} + 4 \text{ muffins (baking)} = 7 \text{ eggs/day} \]

The number of eggs she sells at the farmers' market is the total number of eggs she lays minus the number of eggs she eats:
\[ 16 \text{ eggs} - 7 \text{ eggs} = 9 \text{ eggs/day} \]

**Step 3: Calculate the total revenue from selling the eggs.**

Janet sells each egg for $2. Therefore, the total revenue from selling 9 eggs is:
\[ 9 \text{ eggs} \times 2 \text{ dollars/egg} = 18 \text{ dollars/day} \]

So, the amount Janet makes at the farmers' market every day is:
\[ \boxed{18} \]

Before any training it gets 48.7% of the test questions right. Its reply to the first one lays the problem out in steps, works through them, and boxes 18, which is right. On about half the questions it slips somewhere along the way, and on a few it never gets to a boxed answer at all. That’s the sweet spot for RL: the model can already do the task some of the time, so when we sample several replies, some of them will pass.

Why nudging towards the right replies works

Section titled “Why nudging towards the right replies works”

How should a check, which only says yes or no, change the weights? Let’s shrink the problem as far as it goes. Say the model had to pick one of five answers, and only one is right. Its policy is a softmax over five numbers, πθ=softmax(θ)\pi_\theta = \mathrm{softmax}(\theta), and we’d like to raise the expected reward J(θ)=∑aπθ(a) R(a)J(\theta) = \sum_a \pi_\theta(a)\, R(a).

But a checker has no gradient; you can’t differentiate “the boxed number was 72”. There’s a neat way around that. Because ∇π=π ∇log⁡π\nabla \pi = \pi\, \nabla \log \pi, the gradient of the expected reward becomes an average over the model’s own tries:

∇J=Ea∼πθ[R(a) ∇log⁡πθ(a)].\nabla J = \mathbb{E}_{a \sim \pi_\theta}\big[R(a)\, \nabla \log \pi_\theta(a)\big].

Read it out loud: sample an answer, look at its reward, and push its log-probability up by that much. That’s REINFORCE (Williams, 1992), and everything after it is a refinement of this one line. For the softmax the push is ∇θlog⁡πθ(a)=onehot(a)−πθ\nabla_\theta \log \pi_\theta(a) = \mathrm{onehot}(a) - \pi_\theta, which raises the sampled answer and lowers every answer in proportion to how likely it already was. A reply is the same kind of choice, made one token at a time.

The catch is that word “average”. A reply costs hundreds of tokens to generate, so we only get a handful per question, and a handful of scores is a noisy estimate. If none of them is right, there’s nothing to push up at all. No trick recovers a signal from tries that all failed, which is why RLVR needs questions the model sometimes gets right. But even when some tries do pass, comparing each try against its siblings tames a lot of the noise.

Compare each reply with its group

Section titled “Compare each reply with its group”
Eight replies to one question scored 1 or 0, their mean, and the advantages: each reward minus the mean, divided by the group's standard deviation.reward R of 8 replies to one question10011000mean 0.375advantage A = (R − 0.375) / 0.52+1.21-0.72-0.72+1.21+1.21-0.72-0.72-0.72right replies are pushed up, wrong ones down
Eight replies to one question scored 1 or 0, their mean, and the advantages: each reward minus the mean, divided by the group's standard deviation.reward R of 8 replies to one question10011000mean 0.375advantage A = (R − 0.375) / 0.52+1.21-0.72-0.72+1.21+1.21-0.72-0.72-0.72right replies are pushed up, wrong ones down

Say a reply scores 1. How much should it go up? That depends on the question. On a question the model almost always gets right, one more right reply tells us little, and on one it usually fails, a right reply is exactly what we want more of. So GRPO samples a group of GG replies to the same question and judges each against the others, as in the picture: eight scores, their mean, and how far each one sits from the mean in units of the group’s spread.

A(i)=R(i)−mean⁡jR(j)std⁡jR(j)+10−6.A^{(i)} = \frac{R^{(i)} - \operatorname{mean}_j R^{(j)}}{\operatorname{std}_j R^{(j)} + 10^{-6}}.

Beat your group and you’re pushed up, fall short and you’re pushed down, and dividing by the spread puts easy and hard questions on one scale. When every reply in a group scores the same, all right or all wrong, every advantage is zero and the question teaches nothing this time. (The 10−610^{-6} only keeps us from dividing by zero.)

How does that look on real replies? Below, our untrained model gets eight tries at one training question. Each try is shown by the answer it boxed, or by its last few tokens if it never boxed one.

def record(index, row):
    # A question as a prompt record: the chat, the reference the checker compares with,
    # and the question's index, which reaches the checker as its extra information.
    return {"prompt": messages(row), "data_source": "gsm8k",
            "reward_model": {"ground_truth": reference(row)},
            "extra_info": {"index": index}}




width = PROMPT_TOKENS + NEW_TOKENS
# Full recomputation in the backward pass keeps a step's activations small.
policy = source.model.clone(remat=remat_policy("full"))
# With beta > 0 the objective keeps a frozen copy of the starting weights for the KL.
objective = GRPOObjective(policy, width - 1, variables=source.variables, beta=BETA)
sampling = Sampling(temperature=1.0, **ends)
rollout = SampledRollout(objective, checker, decode=tokenizer.decode, groups=GROUPS,
                         max_new_tokens=NEW_TOKENS, sampling=sampling)
records = tuple(record(index, row) for index, row in enumerate(train))
data = loader(PromptData(source=records, tokenizer=MODEL, max_prompt_len=PROMPT_TOKENS,
                         pad_id=stock.pad_token_id),
              batch=PROMPTS, read="memory", loading=Loading(workers=0))
questions = next(iter(data.train(DataPartition())))
start = SimpleNamespace(variables=source.variables, updates=0)
batch = rollout(start, questions, jax.random.key(SEED))
# The batch packs replies into rows, several to a row where they fit. session_index
# says which reply a token belongs to, and replies q * GROUPS to q * GROUPS + GROUPS - 1
# answer the batch's question q.
def written(s):
    return (batch["session_index"] == s) & (batch["response_mask"] > 0)




sessions = range(PROMPTS * GROUPS)
by_question = np.array([batch["advantages"][written(s)][0] for s in sessions])
by_question = by_question.reshape(PROMPTS, GROUPS)
# The first question whose replies didn't all score the same.
q = next(q for q in range(PROMPTS) if np.ptp(by_question[q]) > 0)
group, advantages = range(q * GROUPS, (q + 1) * GROUPS), by_question[q]
# A finished reply ends with the end-of-turn token, which the checker never sees.
replies = [[token for token in batch["input_ids"][written(s)].tolist()
            if token not in stock.eos_token_ids] for s in group]
info = bytes(np.asarray(questions[INFO_KEY][q], np.uint8)).rstrip(b"\0")
asked = train[json.loads(info)["index"]]
rewards = np.array([checker(tokenizer.decode(reply), PromptRecord("gsm8k", reference(asked)))
                    for reply in replies])
print(asked["question"], "\nreference:", reference(asked))
A math teacher had $100 to buy three different types of calculators. A basic calculator costs $8. A scientific calculator costs twice the price as the basic while a graphing calculator costs thrice the price as the scientific. How much change did she receive after buying those three different types of calculators? 
reference: 28
def answer(reply, shown=6):
    # The tokens from the last "boxed" on, or the reply's last few without one.
    pieces = [tokenizer.decode([token]) for token in reply]
    boxed = [i for i, piece in enumerate(pieces) if "boxed" in piece]
    start = boxed[-1] if boxed else len(pieces) - shown
    return [piece.strip() or "·" for piece in pieces[start:start + shown]]




shown = [answer(reply) for reply in replies]
shade = ["accent_fill" if a > 0 else "warm_fill" for a in advantages]
figures.tokens(shown, fills=[[fill] * len(row) for fill, row in zip(shade, shown)],
               notes=[f"R {r:.0f}  A {a:+.2f}" for r, a in zip(rewards, advantages)])
token rowsboxed{28}\).R 1 A +0.72boxed{28}\)R 1 A +0.72·100=·R 0 A -1.21boxed{0}\).R 0 A -1.21boxed{18}.R 0 A -1.21boxed{28}\).R 1 A +0.72boxed{28}\)R 1 A +0.72boxed{28}\)R 1 A +0.72

Five of the eight replies box 28, the reference, and score 1. Of the other three, one boxes 0, one boxes 18, and one runs out of room in the middle of a sum. With 5 of 8 right the group’s mean is 0.625, so a right reply sits +0.72 above it in units of the spread and a wrong one −1.21 below. Notice the asymmetry: on a question the model mostly gets right, a miss gets pushed down harder than a hit gets pushed up. On a question it mostly fails, it would be the other way round.

From one choice to a whole reply

Section titled “From one choice to a whole reply”
The tokens of one reply: the end of the question, in all ?, with mask 0, and the tokens the policy wrote, 48 + 24 = 72, with mask 1, each carrying the reply's advantage +0.87.tokensmask mA × min00all00?00481+0.87+1+0.87241+0.87=1+0.87721+0.87reward 1 gives A = +0.87 on each written token, 0 on the question

A reply is a sequence of choices, one per token, and its few hundred tokens have to share one reward. Because a reply’s log-probability is the sum of its tokens’ log-probabilities, we give every token the model wrote the same advantage, its reply’s, as in the picture. The model didn’t choose the question, and a mask mm (1 on written tokens, 0 everywhere else) leaves those tokens out. With ℓt=log⁡πθ(yt∣q,y<t)\ell_t = \log \pi_\theta(y_t \mid q, y_{<t}) for one token, Dew minimizes the average over every written token in the batch:

L(θ)=−∑i,tmt(i)A(i)ℓt(i)∑i,tmt(i).L(\theta) = -\frac{\sum_{i,t} m^{(i)}_t A^{(i)} \ell^{(i)}_t}{\sum_{i,t} m^{(i)}_t}.

Does Dew really compute this? We can check on the group above: the advantages straight from the checker’s scores, and LL written out from Qwen’s own next-token probabilities, its gradient against the gradient of Dew’s loss.

# Without the division by the spread, centring on a mean that includes each reply's
# own reward shrinks the expected gradient by (G - 1) / G without turning it;
# dew.rl.rloo_advantage leaves the reply out.
ours = (rewards - rewards.mean()) / (rewards.std(ddof=1) + 1e-6)
np.testing.assert_allclose(advantages, ours, rtol=1e-5, atol=1e-6)
print("the advantages match")
the advantages match
# The rows that hold the group, with the loss kept to its replies.
held = np.isin(batch["session_index"], group)
rows = {name: column[held.any(axis=1)] for name, column in batch.items()}
rows["response_mask"] = rows["response_mask"] * held[held.any(axis=1)]
step = Step(step=jnp.zeros((), jnp.int32), key=jax.random.key(0), ema=source.variables)




def relative_gap(a, b):
    # How far apart two gradients are, as a share of the second one's length.
    a, b = jax.tree.leaves(a), jax.tree.leaves(b)
    gap = sum(float(jnp.sum((x - y) ** 2)) for x, y in zip(a, b))
    return (gap / sum(float(jnp.sum(y ** 2)) for y in b)) ** 0.5
def dews_loss(params, model):
    # Every written token rescored at the starting weights makes each ratio exactly 1,
    # and the policy is still the reference, so neither the clip nor the KL moves it.
    graded = GRPOObjective(model, width - 1, variables=source.variables, beta=BETA)
    scored = rows | {"old_log_probs": graded.packed_log_probs(source.variables, rows)}
    statistics, _ = graded.loss(params, scored, step)
    return statistics.mean()[0]




def our_loss(params, model):
    tokens = jnp.asarray(rows["input_ids"])
    m, a = jnp.asarray(rows["response_mask"]), jnp.asarray(rows["advantages"])
    packing = {"segment_ids": jnp.asarray(rows["text_segment_ids"])[:, :-1],
               "positions": jnp.asarray(rows["text_positions"])[:, :-1]}
    logits = model.apply(params, tokens[:, :-1], **packing).astype(jnp.float32)
    log_probs = jax.nn.log_softmax(logits)
    ell = jnp.take_along_axis(log_probs, tokens[:, 1:, None], axis=-1)[..., 0]
    return -jnp.sum(m[:, 1:] * a[:, 1:] * ell) / jnp.sum(m[:, 1:])




for dtype, bound in ((jnp.bfloat16, 0.02), (jnp.float32, 0.002)):
    model = policy.clone(dtype=dtype)
    gap = relative_gap(jax.grad(dews_loss)(source.variables, model),
                       jax.grad(our_loss)(source.variables, model))
    assert gap < bound
    print(f"{jnp.dtype(dtype).name}: the gradients differ by {gap:.2%} of their length")
bfloat16: the gradients differ by 1.04% of their length
float32: the gradients differ by 0.08% of their length

Both match. In bfloat16, the precision the model computes in, the two gradients differ by 1.04% of their length; run in float32, the same comparison agrees to 0.08%. So the gap is rounding. bfloat16 keeps only two or three significant digits, and Dew’s loss and ours do the same sums in a different order.

Taking more than one step per batch

Section titled “Taking more than one step per batch”
The clipped objective against the ratio for a positive and a negative advantage. Where it is flat, the clipped term gives the token no gradient.clipped0123A > 0: raise π(y), up to ρ = 1 + εratio ρ = π / π_oldobjectiveclippeddual clip0123A < 0: lower π(y), down to ρ = 1 − εratio ρ = π / π_oldobjectiveshaded: flat, so the clipped term gives no gradient
The clipped objective against the ratio for a positive and a negative advantage. Where it is flat, the clipped term gives the token no gradient.clipped0123A > 0: raise π(y), up to ρ = 1 + εratio ρ = π / π_oldobjectiveclippeddual clip0123A < 0: lower π(y), down to ρ = 1 − εratio ρ = π / π_oldobjectiveshaded: flat, so the clipped term gives no gradient

Generating replies is the slow part, so we’d like more than one gradient step per batch, or to train on replies a separate engine wrote while the last update ran. Either way the model has moved since it wrote the batch. Can we still train on it?

PPO (Schulman et al., 2017) weights each token by the ratio ρt=πθ(yt)/πold(yt)\rho_t = \pi_\theta(y_t) / \pi_\text{old}(y_t), how much more or less likely the current model finds it than the model that wrote it, and stops pushing once that ratio has moved far enough the way the advantage asks:

max⁡(−ρtA,  −clip⁡(ρt,0.8,1.2) A).\max\big(-\rho_t A,\; -\operatorname{clip}(\rho_t, 0.8, 1.2)\, A\big).

In the picture, a positive advantage keeps improving the objective until the ratio reaches 1.2. Past that the line is flat, so while the ratio stays above 1.2 the token gets no gradient from the reward term. A negative advantage does the same below 0.8. For negative advantages Dew also caps the term at −3A-3A, the dual clip of Ye et al., 2020, so one token with a runaway ratio can’t take over a step.

ratios = np.array([0.5, 0.9, 1.1, 1.5, 4.0])
for advantage in (1.0, -1.0):
    formula = np.maximum(-ratios * advantage, -np.clip(ratios, 0.8, 1.2) * advantage)
    if advantage < 0:
        formula = np.minimum(formula, -3.0 * advantage)
    terms, _ = clipped_surrogate_terms(jnp.log(ratios)[None], jnp.asarray([advantage]),
                                       jnp.ones((1, ratios.size)))
    np.testing.assert_allclose(terms[0], formula, rtol=1e-6)
    print(f"A = {advantage:+.0f}: terms {np.asarray(terms[0])} at ratios {ratios}")
A = +1: terms [-0.5 -0.9 -1.1 -1.2 -1.2] at ratios [0.5 0.9 1.1 1.5 4. ]
A = -1: terms [0.8 0.9 1.1 1.5 3. ] at ratios [0.5 0.9 1.1 1.5 4. ]

The terms follow the formula. With a positive advantage they stop at −1.2 past a ratio of 1.2, and with a negative one they sit at 0.8 below 0.8 and stop at 3 when the dual clip takes over.

Keeping the model close to where it started

Section titled “Keeping the model close to where it started”
The k3 estimate of the KL divergence against the ratio of the policy to the reference: zero where they agree and rising on both sides, so its gradient pulls the policy back toward the reference.0.250.5124π / π_ref, the policy over the referencek3 per tokenzero where π = π_ref

GRPO’s loss has one more term. While it chases the reward, the model can drift far from where it started and lose what it knew, so GRPO adds β\beta times a per-token estimate of the KL divergence from the starting policy πref\pi_\text{ref}:

k3=πrefπθ−log⁡πrefπθ−1  ≥0.k_3 = \frac{\pi_\text{ref}}{\pi_\theta} - \log \frac{\pi_\text{ref}}{\pi_\theta} - 1 \;\ge 0.

It’s zero when the two policies agree and grows as the model drifts either way, so its gradient points back towards the start, like the ball in the picture. Where the clip has flattened the reward term, this one can still pull, unless it has reached its own cap. Dew caps it at 10, as verl does, so one token the model has nearly ruled out can’t swamp the loss.

policy_probs = np.array([0.1, 0.3, 0.5, 0.9, 0.01])
reference_probs = np.array([0.3, 0.3, 0.25, 0.5, 0.2])
# Dew caps k3 at 10, and the log-ratio inside it at ±20, as verl does.
d = np.clip(np.log(reference_probs / policy_probs), -20, 20)
estimate = k3_kl(jnp.log(policy_probs), jnp.log(reference_probs))
capped = np.minimum(np.exp(d) - d - 1, 10)
np.testing.assert_allclose(estimate, capped, rtol=1e-5, atol=1e-7)
print("formula:", np.exp(d) - d - 1)
print("k3_kl:  ", np.asarray(estimate))
formula: [ 0.901  0.     0.193  0.143 16.004]
k3_kl:   [ 0.901  0.     0.193  0.143 10.   ]

The first four values match the formula. For the last pair the formula gives 16.0, and Dew stops at the cap of 10.

Training it

Section titled “Training it”

Now we put the pieces together. Do right replies get more common? The mean reward is the share of sampled replies that pass the checker, so we can watch it, and also how many replies run out of room before they finish.

optimizer = OptimConfig(learning_rate=LEARNING_RATE, b2=0.99, weight_decay=0.0,
                        clip_grads=1.0)
run = f"runs/10-rlvr/{time.strftime('%Y%m%d-%H%M%S')}"  # a fresh journal per run
trainer = Trainer(objective, optimizer, key=jax.random.key(SEED), rollout=rollout,
                  tracker=LocalTracker(run))
state = trainer.fit(data, steps=STEPS, log_every=1)


with open(f"{run}/scalars.jsonl") as lines:
    logged = [json.loads(line) for line in lines]
logged = [row for row in logged if "rollout/reward/mean" in row["scalars"]]
reward = [row["scalars"]["rollout/reward/mean"] for row in logged]
cut_off = [row["scalars"]["rollout/status/truncated"] for row in logged]
figures.curve({"mean reward": reward, "cut off": cut_off},
              x=[row["step"] for row in logged], x_label="step",
              y_label="share of replies", y_range=(0, 1), smooth=25)
Training CausalTransformer from step 0 to 300: 494.0M parameters, on 1 × NVIDIA A100-SXM4-40GB, batch 16, float32 parameters, bfloat16 compute
13:23:39 WARNING dew.nn.kernels.tuning: 6 kernel launches have no config measured on NVIDIA A100-SXM4-40GB, so they run the nearest measured shape's or their heuristic: measure them on a GPU to yourself with Trainer(tune_kernels=True) or `python tools/kernel_picks.py`
train/mfu is unavailable for GRPOObjective: GRPOObjective needs counts for its auxiliary model forwards
step 1/300  loss 0.04900  actor/pg_clipfrac 8.865e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.835e-04  kl 0  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04900  data_wait_ms 0  grad_norm 0.8760  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 6,020  samples_per_sec 21.26  hlo_flops_util_lower_bound 23.6%  hfu_lower_bound 26.5%  unsized_kernels 291.0  rollout_seconds 27.61  accepted 100.0%  learning_rate 3.000e-06  length/mean 317.2  reward/mean 0.5391  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0
step 2/300  loss 0.02207  actor/pg_clipfrac 7.647e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.411e-04  kl 4.033e-04  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02206  data_wait_ms 0.04483  grad_norm 0.7588  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 30,740  samples_per_sec 4.164  hlo_flops_util_lower_bound 4.6%  hfu_lower_bound 5.2%  unsized_kernels 291.0  rollout_seconds 11.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 306.5  reward/mean 0.5312  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:09:45 left
step 3/300  loss 0.01741  actor/pg_clipfrac 8.513e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.674e-04  kl 6.922e-04  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01739  data_wait_ms 0.02695  grad_norm 0.7917  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,883  samples_per_sec 7.581  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.68  accepted 100.0%  learning_rate 3.000e-06  length/mean 330.4  reward/mean 0.2812  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:16:29 left
step 4/300  loss 0.005420  actor/pg_clipfrac 0.001291  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.390e-04  kl 0.001437  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.005363  data_wait_ms 0.03191  grad_norm 0.9584  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,930  samples_per_sec 7.561  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 332.8  reward/mean 0.4609  status/truncated 7.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:18:35 left
step 5/300  loss 0.01116  actor/pg_clipfrac 9.689e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.173e-04  kl 0.001928  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01108  data_wait_ms 0.02766  grad_norm 0.8565  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,796  samples_per_sec 7.621  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.75  accepted 100.0%  learning_rate 3.000e-06  length/mean 330.6  reward/mean 0.4688  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:19:33 left
step 6/300  loss 0.03677  actor/pg_clipfrac 0.001410  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.609e-04  kl 0.002269  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03668  data_wait_ms 0.02872  grad_norm 0.9139  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,959  samples_per_sec 7.547  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.62  accepted 100.0%  learning_rate 3.000e-06  length/mean 310.2  reward/mean 0.4766  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:19:53 left
step 7/300  loss 0.03142  actor/pg_clipfrac 0.001089  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.354e-04  kl 0.002678  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03131  data_wait_ms 0.03083  grad_norm 0.8310  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,801  samples_per_sec 7.619  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.78  accepted 100.0%  learning_rate 3.000e-06  length/mean 330.0  reward/mean 0.4062  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:20:09 left
step 8/300  loss 0.06012  actor/pg_clipfrac 0.001381  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.434e-04  kl 0.004932  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.05992  data_wait_ms 0.03013  grad_norm 0.9149  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,821  samples_per_sec 7.609  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.62  accepted 100.0%  learning_rate 3.000e-06  length/mean 305.4  reward/mean 0.4609  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:20:09 left
step 9/300  loss 0.02985  actor/pg_clipfrac 0.001225  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.057e-04  kl 0.006408  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02959  data_wait_ms 0.02880  grad_norm 0.9205  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,940  samples_per_sec 7.556  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 287.1  reward/mean 0.5859  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:20:05 left
step 10/300  loss 0.01338  actor/pg_clipfrac 4.460e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.638e-04  kl 0.006723  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01311  data_wait_ms 0.02968  grad_norm 0.7314  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,920  samples_per_sec 7.565  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.76  accepted 100.0%  learning_rate 3.000e-06  length/mean 280.2  reward/mean 0.5156  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:20:02 left
step 11/300  loss 0.01369  actor/pg_clipfrac 0.001008  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.573e-04  kl 0.006913  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01342  data_wait_ms 0.03028  grad_norm 0.8297  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,992  samples_per_sec 7.533  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.74  accepted 100.0%  learning_rate 3.000e-06  length/mean 317.9  reward/mean 0.5000  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:19:56 left
step 12/300  loss 0.01259  actor/pg_clipfrac 0.001457  actor/pg_clipfrac_lower 0  actor/ppo_kl 7.660e-04  kl 0.008666  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01224  data_wait_ms 0.03047  grad_norm 0.8456  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,830  samples_per_sec 7.606  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.81  accepted 100.0%  learning_rate 3.000e-06  length/mean 321.8  reward/mean 0.4766  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:19:50 left
step 13/300  loss -0.01760  actor/pg_clipfrac 8.873e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.609e-04  kl 0.008764  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01796  data_wait_ms 0.02683  grad_norm 0.9171  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,929  samples_per_sec 7.561  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 308.2  reward/mean 0.4609  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:19:38 left
step 14/300  loss 0.04753  actor/pg_clipfrac 0.001059  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.116e-04  kl 0.009662  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04715  data_wait_ms 0.04163  grad_norm 0.9531  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,709  samples_per_sec 7.660  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.75  accepted 100.0%  learning_rate 3.000e-06  length/mean 332.0  reward/mean 0.5312  status/truncated 9.4%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:19:28 left
step 15/300  loss 0.01230  actor/pg_clipfrac 0.001005  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.657e-04  kl 0.01060  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01188  data_wait_ms 0.02915  grad_norm 0.8923  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,899  samples_per_sec 7.574  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.53  accepted 100.0%  learning_rate 3.000e-06  length/mean 287.7  reward/mean 0.5547  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:19:12 left
step 16/300  loss 0.02030  actor/pg_clipfrac 0.001088  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.172e-04  kl 0.01158  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01984  data_wait_ms 0.03099  grad_norm 0.9599  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,838  samples_per_sec 7.602  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 294.5  reward/mean 0.6094  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:18:59 left
step 17/300  loss 0.03310  actor/pg_clipfrac 6.563e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.283e-04  kl 0.01313  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03258  data_wait_ms 0.03645  grad_norm 0.8728  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,764  samples_per_sec 7.636  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.66  accepted 100.0%  learning_rate 3.000e-06  length/mean 297.6  reward/mean 0.5625  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:18:45 left
step 18/300  loss 0.04152  actor/pg_clipfrac 9.872e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.700e-04  kl 0.01299  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04101  data_wait_ms 0.02939  grad_norm 0.9162  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,788  samples_per_sec 7.625  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.58  accepted 100.0%  learning_rate 3.000e-06  length/mean 292.8  reward/mean 0.6328  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:18:30 left
step 19/300  loss 0.03206  actor/pg_clipfrac 0.001491  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.358e-04  kl 0.01266  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03155  data_wait_ms 0.02596  grad_norm 0.8760  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,951  samples_per_sec 7.551  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.61  accepted 100.0%  learning_rate 3.000e-06  length/mean 298.7  reward/mean 0.5859  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:18:14 left
step 20/300  loss 0.05838  actor/pg_clipfrac 0.001109  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.759e-04  kl 0.01508  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.05778  data_wait_ms 0.02834  grad_norm 0.9723  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,992  samples_per_sec 7.533  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.77  accepted 100.0%  learning_rate 3.000e-06  length/mean 309.9  reward/mean 0.5781  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:18:01 left
step 21/300  loss 0.03848  actor/pg_clipfrac 8.324e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.203e-04  kl 0.01473  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03789  data_wait_ms 0.02874  grad_norm 0.8690  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,817  samples_per_sec 7.611  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.81  accepted 100.0%  learning_rate 3.000e-06  length/mean 328.5  reward/mean 0.4297  status/truncated 12.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:17:48 left
step 22/300  loss -0.009164  actor/pg_clipfrac 0.001360  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.879e-04  kl 0.01431  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.009737  data_wait_ms 0.03035  grad_norm 0.9809  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,878  samples_per_sec 7.584  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 287.1  reward/mean 0.5938  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:17:33 left
step 23/300  loss 0.01355  actor/pg_clipfrac 0.001430  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.840e-04  kl 0.01768  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01284  data_wait_ms 0.02942  grad_norm 1.003  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,876  samples_per_sec 7.585  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 306.0  reward/mean 0.4453  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:17:18 left
step 24/300  loss 0.01096  actor/pg_clipfrac 0.001031  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.897e-04  kl 0.01529  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01035  data_wait_ms 0.03085  grad_norm 0.8070  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,815  samples_per_sec 7.612  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 325.9  reward/mean 0.5156  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:17:03 left
step 25/300  loss 0.03389  actor/pg_clipfrac 0.001133  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.501e-04  kl 0.01931  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03312  data_wait_ms 0.02873  grad_norm 0.7742  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,923  samples_per_sec 7.564  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 282.6  reward/mean 0.5938  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:17:19 left
step 26/300  loss 0.007307  actor/pg_clipfrac 8.744e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.253e-04  kl 0.01580  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.006675  data_wait_ms 0.03249  grad_norm 0.7869  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,866  samples_per_sec 7.589  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.74  accepted 100.0%  learning_rate 3.000e-06  length/mean 339.5  reward/mean 0.4219  status/truncated 13.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:17:03 left
step 27/300  loss 0.02238  actor/pg_clipfrac 0.001299  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.650e-04  kl 0.01632  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02172  data_wait_ms 0.03547  grad_norm 0.9200  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,854  samples_per_sec 7.595  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.69  accepted 100.0%  learning_rate 3.000e-06  length/mean 312.6  reward/mean 0.6328  status/truncated 7.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:16:46 left
step 28/300  loss 0.02915  actor/pg_clipfrac 0.001445  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.175e-04  kl 0.01769  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02844  data_wait_ms 0.02699  grad_norm 1.002  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,849  samples_per_sec 7.597  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 297.4  reward/mean 0.5469  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:16:28 left
step 29/300  loss 0.02058  actor/pg_clipfrac 9.422e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.396e-04  kl 0.01793  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01986  data_wait_ms 0.02875  grad_norm 0.8355  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,852  samples_per_sec 7.596  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 306.8  reward/mean 0.5938  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:16:12 left
step 30/300  loss -0.001732  actor/pg_clipfrac 9.642e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.259e-04  kl 0.02149  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.002591  data_wait_ms 0.02936  grad_norm 0.8038  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,833  samples_per_sec 7.604  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 316.0  reward/mean 0.5781  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:15:54 left
step 31/300  loss 0.03919  actor/pg_clipfrac 8.299e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.389e-04  kl 0.01981  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03839  data_wait_ms 0.02756  grad_norm 0.9505  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,580  samples_per_sec 7.720  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 301.3  reward/mean 0.5703  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:15:37 left
step 32/300  loss 0.01442  actor/pg_clipfrac 0.001552  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.945e-04  kl 0.02314  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01349  data_wait_ms 0.02747  grad_norm 1.027  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,827  samples_per_sec 7.607  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.40  accepted 100.0%  learning_rate 3.000e-06  length/mean 261.8  reward/mean 0.6953  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:15:18 left
step 33/300  loss 0.01726  actor/pg_clipfrac 9.562e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.550e-04  kl 0.01861  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01652  data_wait_ms 0.02473  grad_norm 0.9444  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,898  samples_per_sec 7.575  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 302.3  reward/mean 0.5312  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:14:59 left
step 34/300  loss 0.01865  actor/pg_clipfrac 9.082e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.723e-04  kl 0.01736  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01795  data_wait_ms 0.02815  grad_norm 0.6775  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,870  samples_per_sec 7.587  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 309.7  reward/mean 0.5703  status/truncated 9.4%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:14:42 left
step 35/300  loss 0.01365  actor/pg_clipfrac 9.089e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.960e-04  kl 0.01838  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01291  data_wait_ms 0.02972  grad_norm 0.7560  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,644  samples_per_sec 7.690  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.69  accepted 100.0%  learning_rate 3.000e-06  length/mean 309.4  reward/mean 0.5156  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:14:24 left
step 36/300  loss 0.009225  actor/pg_clipfrac 6.379e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.947e-04  kl 0.01984  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.008431  data_wait_ms 0.02953  grad_norm 0.8731  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,729  samples_per_sec 7.651  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.47  accepted 100.0%  learning_rate 3.000e-06  length/mean 257.2  reward/mean 0.6406  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:14:05 left
step 37/300  loss 0.01121  actor/pg_clipfrac 8.051e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.379e-04  kl 0.01981  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01042  data_wait_ms 0.03023  grad_norm 0.8741  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,767  samples_per_sec 7.634  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.55  accepted 100.0%  learning_rate 3.000e-06  length/mean 271.7  reward/mean 0.5234  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:13:46 left
step 38/300  loss 0.007422  actor/pg_clipfrac 0.001060  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.448e-04  kl 0.02062  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.006597  data_wait_ms 0.02867  grad_norm 0.7922  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,732  samples_per_sec 7.650  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.58  accepted 100.0%  learning_rate 3.000e-06  length/mean 302.3  reward/mean 0.6094  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:13:30 left
step 39/300  loss 0.03508  actor/pg_clipfrac 0.001373  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.406e-04  kl 0.02357  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03413  data_wait_ms 0.02820  grad_norm 1.055  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,776  samples_per_sec 7.630  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.55  accepted 100.0%  learning_rate 3.000e-06  length/mean 307.2  reward/mean 0.6797  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:13:11 left
step 40/300  loss 0.04560  actor/pg_clipfrac 9.323e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.554e-04  kl 0.02370  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04465  data_wait_ms 0.02891  grad_norm 0.9673  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,801  samples_per_sec 7.619  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.60  accepted 100.0%  learning_rate 3.000e-06  length/mean 284.9  reward/mean 0.6797  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:12:53 left
step 41/300  loss 0.04736  actor/pg_clipfrac 8.126e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.563e-04  kl 0.02447  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04638  data_wait_ms 0.03407  grad_norm 0.9628  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,893  samples_per_sec 7.577  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.63  accepted 100.0%  learning_rate 3.000e-06  length/mean 307.6  reward/mean 0.5703  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:12:37 left
step 42/300  loss 0.03744  actor/pg_clipfrac 0.001317  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.041e-04  kl 0.02227  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03655  data_wait_ms 0.03049  grad_norm 1.019  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,916  samples_per_sec 7.567  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 302.5  reward/mean 0.6719  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:12:21 left
step 43/300  loss 0.01540  actor/pg_clipfrac 0.001067  actor/pg_clipfrac_lower 0  actor/ppo_kl 8.798e-04  kl 0.02351  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01446  data_wait_ms 0.02859  grad_norm 0.8879  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 17,004  samples_per_sec 7.527  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.74  accepted 100.0%  learning_rate 3.000e-06  length/mean 329.3  reward/mean 0.3828  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:12:04 left
step 44/300  loss 0.03324  actor/pg_clipfrac 0.001192  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.904e-05  kl 0.02289  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03233  data_wait_ms 0.02892  grad_norm 0.9815  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,722  samples_per_sec 7.654  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.83  accepted 100.0%  learning_rate 3.000e-06  length/mean 294.9  reward/mean 0.5078  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:11:47 left
step 45/300  loss 0.03395  actor/pg_clipfrac 0.001327  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.565e-04  kl 0.02370  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03300  data_wait_ms 0.02673  grad_norm 1.121  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 17,011  samples_per_sec 7.525  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.55  accepted 100.0%  learning_rate 3.000e-06  length/mean 288.5  reward/mean 0.5391  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:11:30 left
step 46/300  loss 0.002508  actor/pg_clipfrac 6.388e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.801e-04  kl 0.02228  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.001617  data_wait_ms 0.03016  grad_norm 0.7228  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,891  samples_per_sec 7.578  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.84  accepted 100.0%  learning_rate 3.000e-06  length/mean 342.4  reward/mean 0.4219  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:11:14 left
step 47/300  loss 0.009570  actor/pg_clipfrac 0.001201  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.161e-04  kl 0.02351  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.008630  data_wait_ms 0.03127  grad_norm 0.8907  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,830  samples_per_sec 7.605  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 299.1  reward/mean 0.5156  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:10:58 left
step 48/300  loss 0.02020  actor/pg_clipfrac 0.001225  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.192e-04  kl 0.02344  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01926  data_wait_ms 0.02945  grad_norm 0.9070  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,772  samples_per_sec 7.632  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.66  accepted 100.0%  learning_rate 3.000e-06  length/mean 331.6  reward/mean 0.5078  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:10:41 left
step 49/300  loss -0.003714  actor/pg_clipfrac 9.740e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.464e-04  kl 0.02608  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.004757  data_wait_ms 0.03479  grad_norm 0.8598  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,788  samples_per_sec 7.624  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.60  accepted 100.0%  learning_rate 3.000e-06  length/mean 272.7  reward/mean 0.5547  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:10:22 left
step 50/300  loss 0.04994  actor/pg_clipfrac 0.001117  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.839e-04  kl 0.02618  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04889  data_wait_ms 0.03191  grad_norm 1.016  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,915  samples_per_sec 7.567  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.61  accepted 100.0%  learning_rate 3.000e-06  length/mean 279.7  reward/mean 0.6641  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:10:05 left
step 51/300  loss 0.04122  actor/pg_clipfrac 0.001390  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.868e-04  kl 0.02428  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04025  data_wait_ms 0.02909  grad_norm 1.069  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,854  samples_per_sec 7.594  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.75  accepted 100.0%  learning_rate 3.000e-06  length/mean 281.0  reward/mean 0.6172  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:09:49 left
step 52/300  loss 0.008123  actor/pg_clipfrac 0.001104  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.507e-04  kl 0.02212  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.007238  data_wait_ms 0.03036  grad_norm 0.8919  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,787  samples_per_sec 7.625  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.68  accepted 100.0%  learning_rate 3.000e-06  length/mean 304.4  reward/mean 0.4297  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:09:32 left
step 53/300  loss -2.516e-05  actor/pg_clipfrac 0.001050  actor/pg_clipfrac_lower 0  actor/ppo_kl 8.292e-04  kl 0.02774  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.001135  data_wait_ms 0.02945  grad_norm 0.8513  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,894  samples_per_sec 7.577  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.61  accepted 100.0%  learning_rate 3.000e-06  length/mean 282.9  reward/mean 0.4219  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:09:14 left
step 54/300  loss -0.008305  actor/pg_clipfrac 8.159e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.106e-04  kl 0.02373  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.009255  data_wait_ms 0.03026  grad_norm 0.8918  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,729  samples_per_sec 7.652  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 287.2  reward/mean 0.5078  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:08:58 left
step 55/300  loss 0.04009  actor/pg_clipfrac 0.001117  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.352e-04  kl 0.02403  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03913  data_wait_ms 0.02804  grad_norm 0.9172  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,849  samples_per_sec 7.597  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.55  accepted 100.0%  learning_rate 3.000e-06  length/mean 279.9  reward/mean 0.4297  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:08:43 left
step 56/300  loss -5.107e-04  actor/pg_clipfrac 0.001288  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.720e-04  kl 0.02628  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.001562  data_wait_ms 0.04112  grad_norm 1.007  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,831  samples_per_sec 7.605  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.68  accepted 100.0%  learning_rate 3.000e-06  length/mean 297.3  reward/mean 0.4844  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:08:26 left
step 57/300  loss 0.03093  actor/pg_clipfrac 0.001252  actor/pg_clipfrac_lower 0  actor/ppo_kl -1.559e-06  kl 0.02577  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02990  data_wait_ms 0.02912  grad_norm 0.9336  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,658  samples_per_sec 7.684  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.66  accepted 100.0%  learning_rate 3.000e-06  length/mean 287.0  reward/mean 0.7266  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:08:09 left
step 58/300  loss -0.001583  actor/pg_clipfrac 0.001253  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.354e-04  kl 0.02349  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.002522  data_wait_ms 0.02907  grad_norm 0.9409  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,733  samples_per_sec 7.649  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.49  accepted 100.0%  learning_rate 3.000e-06  length/mean 268.0  reward/mean 0.6484  status/truncated 0.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:07:50 left
step 59/300  loss 0.05552  actor/pg_clipfrac 0.001199  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.843e-04  kl 0.02828  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.05439  data_wait_ms 0.02652  grad_norm 1.080  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,700  samples_per_sec 7.665  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.56  accepted 100.0%  learning_rate 3.000e-06  length/mean 280.1  reward/mean 0.6875  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:07:34 left
step 60/300  loss 0.01578  actor/pg_clipfrac 0.001056  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.722e-04  kl 0.02515  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01477  data_wait_ms 0.02932  grad_norm 1.050  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,866  samples_per_sec 7.589  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.53  accepted 100.0%  learning_rate 3.000e-06  length/mean 266.3  reward/mean 0.6641  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:07:17 left
step 61/300  loss 0.04107  actor/pg_clipfrac 7.296e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.120e-04  kl 0.02517  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04007  data_wait_ms 0.02915  grad_norm 0.8592  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,592  samples_per_sec 7.714  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 299.8  reward/mean 0.5391  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:07:01 left
step 62/300  loss 0.02797  actor/pg_clipfrac 9.945e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.781e-04  kl 0.02773  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02686  data_wait_ms 0.02954  grad_norm 0.9798  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,893  samples_per_sec 7.577  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.42  accepted 100.0%  learning_rate 3.000e-06  length/mean 259.2  reward/mean 0.6797  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:06:43 left
step 63/300  loss 0.02130  actor/pg_clipfrac 0.001463  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.515e-04  kl 0.02461  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02031  data_wait_ms 0.03061  grad_norm 1.098  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,821  samples_per_sec 7.610  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 293.8  reward/mean 0.6094  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:06:27 left
step 64/300  loss 0.01937  actor/pg_clipfrac 9.736e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.369e-04  kl 0.02354  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01843  data_wait_ms 0.03008  grad_norm 0.8787  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,853  samples_per_sec 7.595  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 312.9  reward/mean 0.5938  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:06:10 left
step 65/300  loss 0.01691  actor/pg_clipfrac 0.001629  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.402e-04  kl 0.02363  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01596  data_wait_ms 0.03034  grad_norm 1.117  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,902  samples_per_sec 7.573  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.68  accepted 100.0%  learning_rate 3.000e-06  length/mean 297.4  reward/mean 0.6094  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:05:53 left
step 66/300  loss 0.02780  actor/pg_clipfrac 9.434e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.801e-04  kl 0.02340  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02687  data_wait_ms 0.03575  grad_norm 0.8314  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,980  samples_per_sec 7.538  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.73  accepted 100.0%  learning_rate 3.000e-06  length/mean 356.1  reward/mean 0.4609  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:05:36 left
step 67/300  loss 0.03421  actor/pg_clipfrac 0.001351  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.618e-04  kl 0.02208  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03332  data_wait_ms 0.03138  grad_norm 1.018  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,808  samples_per_sec 7.615  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.81  accepted 100.0%  learning_rate 3.000e-06  length/mean 329.6  reward/mean 0.6406  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:05:19 left
step 68/300  loss 0.04397  actor/pg_clipfrac 9.179e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.481e-04  kl 0.02764  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04287  data_wait_ms 0.02799  grad_norm 0.9388  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,837  samples_per_sec 7.602  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 272.4  reward/mean 0.5859  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:05:03 left
step 69/300  loss -0.01934  actor/pg_clipfrac 0.001022  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.170e-04  kl 0.02699  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.02042  data_wait_ms 0.02963  grad_norm 0.8433  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,806  samples_per_sec 7.617  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 290.6  reward/mean 0.6016  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:04:45 left
step 70/300  loss 0.02141  actor/pg_clipfrac 0.001111  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.160e-04  kl 0.02576  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02037  data_wait_ms 0.02925  grad_norm 0.8886  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,739  samples_per_sec 7.647  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.63  accepted 100.0%  learning_rate 3.000e-06  length/mean 288.4  reward/mean 0.5078  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:04:27 left
step 71/300  loss 0.009144  actor/pg_clipfrac 0.001536  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.283e-04  kl 0.03039  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.007929  data_wait_ms 0.02725  grad_norm 1.145  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,784  samples_per_sec 7.627  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.57  accepted 100.0%  learning_rate 3.000e-06  length/mean 279.7  reward/mean 0.5078  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:04:09 left
step 72/300  loss 0.03301  actor/pg_clipfrac 0.001701  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.501e-04  kl 0.02885  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03185  data_wait_ms 0.02703  grad_norm 1.133  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,787  samples_per_sec 7.625  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.61  accepted 100.0%  learning_rate 3.000e-06  length/mean 303.2  reward/mean 0.6094  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:03:53 left
step 73/300  loss 0.05495  actor/pg_clipfrac 0.001270  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.742e-04  kl 0.03044  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.05373  data_wait_ms 0.02824  grad_norm 1.062  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,839  samples_per_sec 7.601  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.62  accepted 100.0%  learning_rate 3.000e-06  length/mean 295.3  reward/mean 0.6797  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:03:36 left
step 74/300  loss 0.02131  actor/pg_clipfrac 9.426e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.662e-04  kl 0.02628  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02025  data_wait_ms 0.02996  grad_norm 0.8727  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,902  samples_per_sec 7.573  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 298.4  reward/mean 0.5391  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:03:18 left
step 75/300  loss 0.01631  actor/pg_clipfrac 9.216e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.299e-04  kl 0.02638  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01525  data_wait_ms 0.02799  grad_norm 0.7697  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,875  samples_per_sec 7.585  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.73  accepted 100.0%  learning_rate 3.000e-06  length/mean 305.2  reward/mean 0.5078  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:03:02 left
step 76/300  loss -0.006089  actor/pg_clipfrac 8.350e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.837e-04  kl 0.02671  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.007157  data_wait_ms 0.02879  grad_norm 0.7888  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,710  samples_per_sec 7.660  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 299.4  reward/mean 0.4453  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:02:46 left
step 77/300  loss 0.008445  actor/pg_clipfrac 0.001300  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.273e-04  kl 0.03119  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.007197  data_wait_ms 0.02901  grad_norm 0.9928  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,820  samples_per_sec 7.610  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.54  accepted 100.0%  learning_rate 3.000e-06  length/mean 276.5  reward/mean 0.5156  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:02:27 left
step 78/300  loss 0.03213  actor/pg_clipfrac 0.001035  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.254e-04  kl 0.03007  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03092  data_wait_ms 0.02718  grad_norm 0.8528  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,849  samples_per_sec 7.597  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 317.2  reward/mean 0.4844  status/truncated 7.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:02:11 left
step 79/300  loss -0.02276  actor/pg_clipfrac 0.001315  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.735e-04  kl 0.03048  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.02398  data_wait_ms 0.04274  grad_norm 0.9934  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,733  samples_per_sec 7.650  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.68  accepted 100.0%  learning_rate 3.000e-06  length/mean 297.0  reward/mean 0.4766  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:01:55 left
step 80/300  loss 0.02553  actor/pg_clipfrac 0.001107  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.315e-04  kl 0.03401  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02417  data_wait_ms 0.02952  grad_norm 1.052  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,840  samples_per_sec 7.601  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.56  accepted 100.0%  learning_rate 3.000e-06  length/mean 296.3  reward/mean 0.5625  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:01:37 left
step 81/300  loss 0.03646  actor/pg_clipfrac 0.001516  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.862e-04  kl 0.03308  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03514  data_wait_ms 0.03616  grad_norm 1.142  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,892  samples_per_sec 7.578  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 293.7  reward/mean 0.6484  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:01:22 left
step 82/300  loss 0.05609  actor/pg_clipfrac 9.524e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.161e-04  kl 0.02834  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.05496  data_wait_ms 0.02643  grad_norm 0.8136  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,979  samples_per_sec 7.539  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 336.3  reward/mean 0.5312  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:01:06 left
step 83/300  loss 0.02239  actor/pg_clipfrac 0.001395  actor/pg_clipfrac_lower 0  actor/ppo_kl 7.260e-04  kl 0.03041  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02117  data_wait_ms 0.03122  grad_norm 0.9071  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,833  samples_per_sec 7.604  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.81  accepted 100.0%  learning_rate 3.000e-06  length/mean 313.7  reward/mean 0.5078  status/truncated 10.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:00:52 left
step 84/300  loss 0.03224  actor/pg_clipfrac 7.984e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 7.366e-05  kl 0.03197  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03096  data_wait_ms 0.03007  grad_norm 0.7663  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,895  samples_per_sec 7.576  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 293.5  reward/mean 0.5938  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:00:35 left
step 85/300  loss 0.01398  actor/pg_clipfrac 0.001425  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.070e-04  kl 0.03067  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01275  data_wait_ms 0.02831  grad_norm 1.077  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,890  samples_per_sec 7.578  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 312.4  reward/mean 0.5625  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:00:21 left
step 86/300  loss 0.03982  actor/pg_clipfrac 8.008e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.669e-04  kl 0.03365  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03848  data_wait_ms 0.02832  grad_norm 0.9272  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,837  samples_per_sec 7.602  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 292.7  reward/mean 0.5781  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  1:00:04 left
step 87/300  loss 0.02345  actor/pg_clipfrac 8.913e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.793e-04  kl 0.03097  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02221  data_wait_ms 0.02607  grad_norm 0.7835  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,800  samples_per_sec 7.619  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.66  accepted 100.0%  learning_rate 3.000e-06  length/mean 315.5  reward/mean 0.5547  status/truncated 7.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:59:48 left
step 88/300  loss 0.002378  actor/pg_clipfrac 0.001566  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.988e-04  kl 0.03236  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.001084  data_wait_ms 0.02872  grad_norm 0.9604  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,931  samples_per_sec 7.560  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.62  accepted 100.0%  learning_rate 3.000e-06  length/mean 294.3  reward/mean 0.5156  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:59:30 left
step 89/300  loss 0.04370  actor/pg_clipfrac 0.001054  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.523e-04  kl 0.03457  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04231  data_wait_ms 0.02972  grad_norm 0.8296  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,940  samples_per_sec 7.556  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.75  accepted 100.0%  learning_rate 3.000e-06  length/mean 318.6  reward/mean 0.5547  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:59:14 left
step 90/300  loss 0.01626  actor/pg_clipfrac 8.896e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.474e-04  kl 0.03420  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01489  data_wait_ms 0.02986  grad_norm 0.8732  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,896  samples_per_sec 7.576  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.76  accepted 100.0%  learning_rate 3.000e-06  length/mean 324.9  reward/mean 0.5078  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:58:56 left
step 91/300  loss 0.03597  actor/pg_clipfrac 0.001124  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.220e-04  kl 0.03541  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03455  data_wait_ms 0.02869  grad_norm 1.020  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,593  samples_per_sec 7.714  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 299.0  reward/mean 0.6016  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:58:40 left
step 92/300  loss 0.01897  actor/pg_clipfrac 0.001113  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.885e-04  kl 0.03841  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01743  data_wait_ms 0.03457  grad_norm 0.9945  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,859  samples_per_sec 7.592  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.41  accepted 100.0%  learning_rate 3.000e-06  length/mean 252.7  reward/mean 0.7500  status/truncated 0.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:58:21 left
step 93/300  loss 0.005905  actor/pg_clipfrac 0.001363  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.233e-04  kl 0.03239  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.004610  data_wait_ms 0.02473  grad_norm 0.8267  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,797  samples_per_sec 7.621  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.68  accepted 100.0%  learning_rate 3.000e-06  length/mean 298.1  reward/mean 0.6797  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:58:05 left
step 94/300  loss 0.05002  actor/pg_clipfrac 0.001307  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.100e-04  kl 0.03793  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04851  data_wait_ms 0.02897  grad_norm 1.054  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,708  samples_per_sec 7.661  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.62  accepted 100.0%  learning_rate 3.000e-06  length/mean 292.8  reward/mean 0.4922  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:57:49 left
step 95/300  loss 0.001732  actor/pg_clipfrac 6.678e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.520e-04  kl 0.03624  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 2.827e-04  data_wait_ms 0.02572  grad_norm 0.9160  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,843  samples_per_sec 7.600  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.53  accepted 100.0%  learning_rate 3.000e-06  length/mean 280.8  reward/mean 0.6875  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:57:31 left
step 96/300  loss 0.06257  actor/pg_clipfrac 5.342e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.393e-04  kl 0.03348  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.06123  data_wait_ms 0.02659  grad_norm 0.8479  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,868  samples_per_sec 7.588  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 307.1  reward/mean 0.4922  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:57:15 left
step 97/300  loss 0.03013  actor/pg_clipfrac 0.001016  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.802e-04  kl 0.03910  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02856  data_wait_ms 0.02936  grad_norm 0.8575  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,807  samples_per_sec 7.616  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 307.5  reward/mean 0.6172  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:56:58 left
step 98/300  loss 0.03482  actor/pg_clipfrac 0.001094  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.926e-04  kl 0.03642  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03336  data_wait_ms 0.02860  grad_norm 0.9005  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,759  samples_per_sec 7.638  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 285.6  reward/mean 0.6172  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:56:41 left
step 99/300  loss 0.01115  actor/pg_clipfrac 0.001287  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.477e-04  kl 0.04251  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.009454  data_wait_ms 0.03071  grad_norm 1.119  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,731  samples_per_sec 7.650  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.59  accepted 100.0%  learning_rate 3.000e-06  length/mean 273.2  reward/mean 0.5703  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:56:23 left
step 100/300  loss 0.03760  actor/pg_clipfrac 9.554e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.887e-04  kl 0.03808  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03608  data_wait_ms 0.02935  grad_norm 0.9070  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,810  samples_per_sec 7.615  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.56  accepted 100.0%  learning_rate 3.000e-06  length/mean 286.2  reward/mean 0.6641  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:56:06 left
step 101/300  loss 0.01931  actor/pg_clipfrac 9.855e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.730e-04  kl 0.04336  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01758  data_wait_ms 0.04100  grad_norm 0.7835  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,700  samples_per_sec 7.665  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 285.4  reward/mean 0.4844  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:55:49 left
step 102/300  loss 0.03072  actor/pg_clipfrac 6.786e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.290e-04  kl 0.03679  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02925  data_wait_ms 0.02989  grad_norm 0.9237  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,710  samples_per_sec 7.660  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.53  accepted 100.0%  learning_rate 3.000e-06  length/mean 264.8  reward/mean 0.5938  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:55:31 left
step 103/300  loss -5.605e-04  actor/pg_clipfrac 8.467e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.042e-04  kl 0.04007  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.002163  data_wait_ms 0.02865  grad_norm 0.8375  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 17,064  samples_per_sec 7.501  hlo_flops_util_lower_bound 8.3%  hfu_lower_bound 9.3%  unsized_kernels 291.0  rollout_seconds 14.54  accepted 100.0%  learning_rate 3.000e-06  length/mean 286.0  reward/mean 0.5703  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:55:14 left
step 104/300  loss 0.01149  actor/pg_clipfrac 8.491e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.716e-04  kl 0.03992  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.009897  data_wait_ms 0.02880  grad_norm 0.7084  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,728  samples_per_sec 7.652  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.89  accepted 100.0%  learning_rate 3.000e-06  length/mean 294.4  reward/mean 0.6250  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:54:59 left
step 105/300  loss 0.01465  actor/pg_clipfrac 0.001031  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.310e-04  kl 0.03551  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01323  data_wait_ms 0.02989  grad_norm 0.8812  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,693  samples_per_sec 7.668  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.56  accepted 100.0%  learning_rate 3.000e-06  length/mean 272.8  reward/mean 0.5625  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:54:41 left
step 106/300  loss -0.01678  actor/pg_clipfrac 8.046e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.778e-04  kl 0.04259  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01848  data_wait_ms 0.02924  grad_norm 0.9214  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,903  samples_per_sec 7.573  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.52  accepted 100.0%  learning_rate 3.000e-06  length/mean 271.9  reward/mean 0.6016  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:54:22 left
step 107/300  loss 0.04631  actor/pg_clipfrac 0.001044  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.989e-04  kl 0.03404  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04495  data_wait_ms 0.03031  grad_norm 0.9140  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,841  samples_per_sec 7.601  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.74  accepted 100.0%  learning_rate 3.000e-06  length/mean 314.3  reward/mean 0.4375  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:54:05 left
step 108/300  loss 0.01513  actor/pg_clipfrac 0.001375  actor/pg_clipfrac_lower 0  actor/ppo_kl 7.150e-04  kl 0.03715  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01365  data_wait_ms 0.02708  grad_norm 0.9893  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,688  samples_per_sec 7.670  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 267.0  reward/mean 0.6016  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:53:48 left
step 109/300  loss -0.01706  actor/pg_clipfrac 0.001225  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.849e-04  kl 0.04175  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01873  data_wait_ms 0.03474  grad_norm 1.022  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,870  samples_per_sec 7.587  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.52  accepted 100.0%  learning_rate 3.000e-06  length/mean 280.7  reward/mean 0.5781  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:53:30 left
step 110/300  loss 0.02759  actor/pg_clipfrac 7.522e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.828e-04  kl 0.03876  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02604  data_wait_ms 0.02882  grad_norm 0.8034  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,840  samples_per_sec 7.601  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 322.0  reward/mean 0.5312  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:53:13 left
step 111/300  loss -0.003014  actor/pg_clipfrac 9.739e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.512e-04  kl 0.03725  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.004504  data_wait_ms 0.02833  grad_norm 0.8076  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,903  samples_per_sec 7.573  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 320.9  reward/mean 0.5156  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:52:57 left
step 112/300  loss 0.01093  actor/pg_clipfrac 6.596e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.235e-04  kl 0.03559  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.009511  data_wait_ms 0.02886  grad_norm 0.9053  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,865  samples_per_sec 7.590  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.73  accepted 100.0%  learning_rate 3.000e-06  length/mean 296.1  reward/mean 0.5859  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:52:40 left
step 113/300  loss 0.01213  actor/pg_clipfrac 9.549e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.141e-04  kl 0.03337  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01079  data_wait_ms 0.03027  grad_norm 0.8458  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,875  samples_per_sec 7.585  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 294.5  reward/mean 0.5625  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:52:22 left
step 114/300  loss 0.006066  actor/pg_clipfrac 9.373e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.294e-04  kl 0.03623  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.004617  data_wait_ms 0.03500  grad_norm 0.8470  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,865  samples_per_sec 7.590  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 316.7  reward/mean 0.5703  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:52:05 left
step 115/300  loss 0.005725  actor/pg_clipfrac 6.367e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.778e-04  kl 0.03737  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.004230  data_wait_ms 0.03152  grad_norm 0.7563  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,878  samples_per_sec 7.584  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 294.5  reward/mean 0.6250  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:51:51 left
step 116/300  loss 0.02555  actor/pg_clipfrac 9.794e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.587e-04  kl 0.03331  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02422  data_wait_ms 0.02907  grad_norm 0.8451  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,867  samples_per_sec 7.589  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 327.1  reward/mean 0.4219  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:51:34 left
step 117/300  loss 0.01473  actor/pg_clipfrac 9.603e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.955e-04  kl 0.03677  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01326  data_wait_ms 0.02984  grad_norm 1.017  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,728  samples_per_sec 7.652  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 309.1  reward/mean 0.5156  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:51:18 left
step 118/300  loss 0.01831  actor/pg_clipfrac 0.001171  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.840e-04  kl 0.03433  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01693  data_wait_ms 0.02909  grad_norm 0.9090  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,799  samples_per_sec 7.619  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.56  accepted 100.0%  learning_rate 3.000e-06  length/mean 313.5  reward/mean 0.5078  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:51:01 left
step 119/300  loss 0.01449  actor/pg_clipfrac 9.967e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.324e-04  kl 0.03419  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01312  data_wait_ms 0.02883  grad_norm 0.9181  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,792  samples_per_sec 7.623  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.63  accepted 100.0%  learning_rate 3.000e-06  length/mean 313.5  reward/mean 0.4844  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:50:44 left
step 120/300  loss -0.01264  actor/pg_clipfrac 5.348e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.309e-04  kl 0.03497  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01404  data_wait_ms 0.03756  grad_norm 0.7642  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,889  samples_per_sec 7.579  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.62  accepted 100.0%  learning_rate 3.000e-06  length/mean 292.2  reward/mean 0.5156  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:50:26 left
step 121/300  loss 0.009340  actor/pg_clipfrac 8.865e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.044e-04  kl 0.02858  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.008197  data_wait_ms 0.02882  grad_norm 0.8394  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,859  samples_per_sec 7.592  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 299.6  reward/mean 0.6562  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:50:10 left
step 122/300  loss -0.008251  actor/pg_clipfrac 8.648e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.219e-04  kl 0.03407  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.009614  data_wait_ms 0.02872  grad_norm 0.9533  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,783  samples_per_sec 7.627  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.69  accepted 100.0%  learning_rate 3.000e-06  length/mean 289.1  reward/mean 0.6484  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:49:54 left
step 123/300  loss 0.01398  actor/pg_clipfrac 9.181e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.127e-04  kl 0.03567  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01255  data_wait_ms 0.03327  grad_norm 0.8040  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,867  samples_per_sec 7.589  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.62  accepted 100.0%  learning_rate 3.000e-06  length/mean 306.3  reward/mean 0.6562  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:49:38 left
step 124/300  loss 0.001212  actor/pg_clipfrac 0.001289  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.195e-04  kl 0.03432  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -1.607e-04  data_wait_ms 0.02942  grad_norm 0.8746  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,864  samples_per_sec 7.590  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 327.3  reward/mean 0.4219  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:49:21 left
step 125/300  loss 0.006311  actor/pg_clipfrac 6.192e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.247e-04  kl 0.03110  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.005067  data_wait_ms 0.04379  grad_norm 0.6492  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 17,019  samples_per_sec 7.521  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 315.4  reward/mean 0.6484  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:49:06 left
step 126/300  loss 0.01541  actor/pg_clipfrac 5.165e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.162e-04  kl 0.03190  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01413  data_wait_ms 0.02638  grad_norm 0.6892  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,833  samples_per_sec 7.604  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.85  accepted 100.0%  learning_rate 3.000e-06  length/mean 317.7  reward/mean 0.4609  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:48:51 left
step 127/300  loss 7.759e-04  actor/pg_clipfrac 0.001079  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.983e-04  kl 0.03428  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -5.954e-04  data_wait_ms 0.03764  grad_norm 1.007  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,896  samples_per_sec 7.576  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 311.3  reward/mean 0.7188  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:48:33 left
step 128/300  loss 0.01787  actor/pg_clipfrac 5.744e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.512e-04  kl 0.03395  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01651  data_wait_ms 0.02687  grad_norm 0.6719  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,933  samples_per_sec 7.559  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.73  accepted 100.0%  learning_rate 3.000e-06  length/mean 340.0  reward/mean 0.4141  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:48:17 left
step 129/300  loss 0.04351  actor/pg_clipfrac 0.001216  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.068e-04  kl 0.03199  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04223  data_wait_ms 0.02940  grad_norm 0.9136  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,824  samples_per_sec 7.608  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.76  accepted 100.0%  learning_rate 3.000e-06  length/mean 302.0  reward/mean 0.7031  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:48:02 left
step 130/300  loss 0.03235  actor/pg_clipfrac 0.001014  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.069e-04  kl 0.03282  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03104  data_wait_ms 0.03183  grad_norm 0.8474  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,936  samples_per_sec 7.558  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 331.3  reward/mean 0.6406  status/truncated 10.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:47:45 left
step 131/300  loss 0.01516  actor/pg_clipfrac 8.215e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.496e-04  kl 0.03271  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01385  data_wait_ms 0.03080  grad_norm 0.7044  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,839  samples_per_sec 7.601  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.77  accepted 100.0%  learning_rate 3.000e-06  length/mean 332.8  reward/mean 0.4844  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:47:29 left
step 132/300  loss -0.002860  actor/pg_clipfrac 9.597e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.919e-04  kl 0.03246  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.004158  data_wait_ms 0.02932  grad_norm 0.7960  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,773  samples_per_sec 7.631  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 293.1  reward/mean 0.6328  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:47:13 left
step 133/300  loss 0.01401  actor/pg_clipfrac 0.001116  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.072e-04  kl 0.03015  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01281  data_wait_ms 0.03607  grad_norm 0.8613  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,797  samples_per_sec 7.620  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.61  accepted 100.0%  learning_rate 3.000e-06  length/mean 308.0  reward/mean 0.6797  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:46:55 left
step 134/300  loss 0.01154  actor/pg_clipfrac 0.001182  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.806e-04  kl 0.03400  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01018  data_wait_ms 0.02861  grad_norm 0.9382  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,724  samples_per_sec 7.654  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.63  accepted 100.0%  learning_rate 3.000e-06  length/mean 284.3  reward/mean 0.6328  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:46:38 left
step 135/300  loss 0.001762  actor/pg_clipfrac 8.865e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.097e-04  kl 0.03155  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 5.000e-04  data_wait_ms 0.02752  grad_norm 0.7759  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,866  samples_per_sec 7.589  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.56  accepted 100.0%  learning_rate 3.000e-06  length/mean 299.6  reward/mean 0.6797  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:46:20 left
step 136/300  loss -4.201e-04  actor/pg_clipfrac 0.001058  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.731e-04  kl 0.03092  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.001657  data_wait_ms 0.02882  grad_norm 0.8546  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,837  samples_per_sec 7.602  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 310.0  reward/mean 0.6172  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:46:03 left
step 137/300  loss 0.02486  actor/pg_clipfrac 6.518e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.303e-04  kl 0.03059  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02363  data_wait_ms 0.02939  grad_norm 0.6901  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,817  samples_per_sec 7.611  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 311.7  reward/mean 0.6641  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:45:46 left
step 138/300  loss 0.01181  actor/pg_clipfrac 3.746e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.787e-04  kl 0.03178  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01053  data_wait_ms 0.02851  grad_norm 0.6142  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,727  samples_per_sec 7.653  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 292.0  reward/mean 0.6328  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:45:29 left
step 139/300  loss 0.06058  actor/pg_clipfrac 7.822e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.223e-04  kl 0.03011  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.05937  data_wait_ms 0.02807  grad_norm 0.9244  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,876  samples_per_sec 7.585  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.56  accepted 100.0%  learning_rate 3.000e-06  length/mean 279.7  reward/mean 0.8359  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:45:11 left
step 140/300  loss 0.03149  actor/pg_clipfrac 0.001139  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.961e-04  kl 0.03189  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03022  data_wait_ms 0.02932  grad_norm 0.8202  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,895  samples_per_sec 7.576  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 329.2  reward/mean 0.6406  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:44:54 left
step 141/300  loss 0.02483  actor/pg_clipfrac 9.747e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.742e-04  kl 0.03184  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02355  data_wait_ms 0.02915  grad_norm 0.8541  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,966  samples_per_sec 7.544  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.73  accepted 100.0%  learning_rate 3.000e-06  length/mean 312.6  reward/mean 0.5312  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:44:38 left
step 142/300  loss 9.136e-04  actor/pg_clipfrac 9.265e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.288e-04  kl 0.03039  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -3.022e-04  data_wait_ms 0.02876  grad_norm 0.8416  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,771  samples_per_sec 7.632  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.80  accepted 100.0%  learning_rate 3.000e-06  length/mean 320.4  reward/mean 0.6406  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:44:23 left
step 143/300  loss 0.01263  actor/pg_clipfrac 0.001016  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.671e-04  kl 0.03239  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01134  data_wait_ms 0.03010  grad_norm 0.8473  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,887  samples_per_sec 7.580  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.61  accepted 100.0%  learning_rate 3.000e-06  length/mean 276.8  reward/mean 0.7266  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:44:06 left
step 144/300  loss 0.01664  actor/pg_clipfrac 9.993e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.866e-04  kl 0.03590  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01520  data_wait_ms 0.02936  grad_norm 0.9136  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,871  samples_per_sec 7.587  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 297.1  reward/mean 0.5234  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:43:49 left
step 145/300  loss 0.02425  actor/pg_clipfrac 0.001343  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.612e-05  kl 0.02996  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02306  data_wait_ms 0.02856  grad_norm 0.9572  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,662  samples_per_sec 7.682  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 302.5  reward/mean 0.6875  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:43:32 left
step 146/300  loss 0.03014  actor/pg_clipfrac 0.001132  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.442e-04  kl 0.03490  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02874  data_wait_ms 0.03128  grad_norm 0.9996  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,878  samples_per_sec 7.584  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.49  accepted 100.0%  learning_rate 3.000e-06  length/mean 283.0  reward/mean 0.5781  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:43:14 left
step 147/300  loss 0.003881  actor/pg_clipfrac 7.247e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.234e-04  kl 0.03207  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.002598  data_wait_ms 0.02991  grad_norm 0.7669  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,773  samples_per_sec 7.631  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 301.9  reward/mean 0.5938  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:42:58 left
step 148/300  loss 0.03703  actor/pg_clipfrac 3.898e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.621e-04  kl 0.03746  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03553  data_wait_ms 0.02702  grad_norm 0.7538  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,824  samples_per_sec 7.608  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.60  accepted 100.0%  learning_rate 3.000e-06  length/mean 280.6  reward/mean 0.7266  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:42:40 left
step 149/300  loss 0.02579  actor/pg_clipfrac 5.408e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.927e-04  kl 0.03662  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02433  data_wait_ms 0.02753  grad_norm 0.6657  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,790  samples_per_sec 7.624  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.66  accepted 100.0%  learning_rate 3.000e-06  length/mean 303.4  reward/mean 0.5781  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:42:22 left
step 150/300  loss -0.001863  actor/pg_clipfrac 0.001052  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.748e-04  kl 0.03523  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.003272  data_wait_ms 0.02979  grad_norm 0.8874  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,880  samples_per_sec 7.583  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.62  accepted 100.0%  learning_rate 3.000e-06  length/mean 289.6  reward/mean 0.5547  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:42:05 left
step 151/300  loss -0.01849  actor/pg_clipfrac 7.271e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.912e-04  kl 0.03468  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01987  data_wait_ms 0.02754  grad_norm 0.7906  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,696  samples_per_sec 7.667  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 322.4  reward/mean 0.5000  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:41:48 left
step 152/300  loss 0.04917  actor/pg_clipfrac 0.001222  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.676e-04  kl 0.03496  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04777  data_wait_ms 0.02967  grad_norm 1.009  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,908  samples_per_sec 7.570  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.53  accepted 100.0%  learning_rate 3.000e-06  length/mean 300.5  reward/mean 0.6797  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:41:30 left
step 153/300  loss 0.006565  actor/pg_clipfrac 8.243e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.685e-04  kl 0.03230  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.005273  data_wait_ms 0.02951  grad_norm 0.8148  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,901  samples_per_sec 7.573  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.74  accepted 100.0%  learning_rate 3.000e-06  length/mean 322.2  reward/mean 0.5156  status/truncated 7.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:41:13 left
step 154/300  loss 0.02885  actor/pg_clipfrac 8.602e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.583e-04  kl 0.03404  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02749  data_wait_ms 0.04395  grad_norm 0.7969  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,841  samples_per_sec 7.600  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.73  accepted 100.0%  learning_rate 3.000e-06  length/mean 299.7  reward/mean 0.6797  status/truncated 7.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:40:56 left
step 155/300  loss 0.004717  actor/pg_clipfrac 0.001171  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.264e-04  kl 0.03438  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.003342  data_wait_ms 0.03052  grad_norm 0.9770  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,688  samples_per_sec 7.670  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 340.2  reward/mean 0.5469  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:40:40 left
step 156/300  loss 0.02501  actor/pg_clipfrac 0.001385  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.253e-04  kl 0.03679  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02354  data_wait_ms 0.02853  grad_norm 0.9695  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,747  samples_per_sec 7.643  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.52  accepted 100.0%  learning_rate 3.000e-06  length/mean 298.9  reward/mean 0.7344  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:40:22 left
step 157/300  loss 0.04286  actor/pg_clipfrac 6.942e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.189e-04  kl 0.03830  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04132  data_wait_ms 0.02737  grad_norm 0.9254  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,892  samples_per_sec 7.578  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.58  accepted 100.0%  learning_rate 3.000e-06  length/mean 292.6  reward/mean 0.6484  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:40:05 left
step 158/300  loss 0.01088  actor/pg_clipfrac 7.098e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.867e-04  kl 0.03324  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.009548  data_wait_ms 0.03369  grad_norm 0.7805  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,883  samples_per_sec 7.581  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 308.2  reward/mean 0.6172  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:39:49 left
step 159/300  loss 0.01802  actor/pg_clipfrac 8.184e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.008e-04  kl 0.03662  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01656  data_wait_ms 0.02755  grad_norm 0.8714  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,872  samples_per_sec 7.587  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 305.5  reward/mean 0.5312  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:39:33 left
step 160/300  loss 0.02509  actor/pg_clipfrac 8.915e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.641e-04  kl 0.03703  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02361  data_wait_ms 0.03016  grad_norm 0.9359  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,811  samples_per_sec 7.614  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 324.3  reward/mean 0.5547  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:39:16 left
step 161/300  loss 0.001419  actor/pg_clipfrac 0.001246  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.836e-04  kl 0.03393  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 6.208e-05  data_wait_ms 0.03050  grad_norm 1.011  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,960  samples_per_sec 7.547  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 307.2  reward/mean 0.5312  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:38:59 left
step 162/300  loss 0.007332  actor/pg_clipfrac 0.001176  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.181e-04  kl 0.03501  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.005932  data_wait_ms 0.03064  grad_norm 0.8917  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,803  samples_per_sec 7.618  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.79  accepted 100.0%  learning_rate 3.000e-06  length/mean 292.2  reward/mean 0.5234  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:38:44 left
step 163/300  loss 0.01047  actor/pg_clipfrac 9.063e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.827e-04  kl 0.03371  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.009123  data_wait_ms 0.03039  grad_norm 0.8353  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,843  samples_per_sec 7.599  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 284.5  reward/mean 0.6953  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:38:26 left
step 164/300  loss 0.01664  actor/pg_clipfrac 7.572e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.224e-04  kl 0.03752  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01514  data_wait_ms 0.03046  grad_norm 0.9424  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,834  samples_per_sec 7.604  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 288.9  reward/mean 0.6094  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:38:09 left
step 165/300  loss 0.01223  actor/pg_clipfrac 5.191e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.877e-04  kl 0.03553  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01081  data_wait_ms 0.02652  grad_norm 0.7547  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,885  samples_per_sec 7.581  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 316.0  reward/mean 0.4766  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:37:52 left
step 166/300  loss 0.02004  actor/pg_clipfrac 7.795e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.580e-04  kl 0.03722  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01855  data_wait_ms 0.03018  grad_norm 0.7788  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,839  samples_per_sec 7.602  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 310.7  reward/mean 0.6016  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:37:35 left
step 167/300  loss -0.007141  actor/pg_clipfrac 8.137e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.047e-04  kl 0.03468  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.008529  data_wait_ms 0.02978  grad_norm 0.9491  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,791  samples_per_sec 7.623  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 326.4  reward/mean 0.6250  status/truncated 7.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:37:18 left
step 168/300  loss 0.02497  actor/pg_clipfrac 0.001184  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.037e-04  kl 0.04003  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02337  data_wait_ms 0.04429  grad_norm 1.076  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,887  samples_per_sec 7.580  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.62  accepted 100.0%  learning_rate 3.000e-06  length/mean 290.4  reward/mean 0.5703  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:37:01 left
step 169/300  loss 0.02369  actor/pg_clipfrac 9.761e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.501e-04  kl 0.03488  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02230  data_wait_ms 0.02898  grad_norm 0.9184  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,629  samples_per_sec 7.697  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 304.1  reward/mean 0.7578  status/truncated 9.4%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:36:46 left
step 170/300  loss 0.03224  actor/pg_clipfrac 9.059e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.790e-04  kl 0.04079  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03061  data_wait_ms 0.02683  grad_norm 1.062  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,735  samples_per_sec 7.649  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.46  accepted 100.0%  learning_rate 3.000e-06  length/mean 267.4  reward/mean 0.7031  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:36:27 left
step 171/300  loss 0.03233  actor/pg_clipfrac 0.001712  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.774e-04  kl 0.03688  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03086  data_wait_ms 0.03011  grad_norm 1.064  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,816  samples_per_sec 7.612  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.57  accepted 100.0%  learning_rate 3.000e-06  length/mean 283.0  reward/mean 0.7109  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:36:10 left
step 172/300  loss 0.005895  actor/pg_clipfrac 0.001019  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.672e-04  kl 0.03792  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.004378  data_wait_ms 0.02967  grad_norm 0.9516  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,862  samples_per_sec 7.591  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 291.3  reward/mean 0.5547  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:35:53 left
step 173/300  loss -9.085e-04  actor/pg_clipfrac 9.115e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.237e-04  kl 0.03349  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.002248  data_wait_ms 0.02762  grad_norm 0.9024  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,704  samples_per_sec 7.663  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 282.8  reward/mean 0.6484  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:35:37 left
step 174/300  loss -0.02489  actor/pg_clipfrac 0.001022  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.556e-04  kl 0.03894  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.02644  data_wait_ms 0.02922  grad_norm 0.8840  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,639  samples_per_sec 7.693  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.53  accepted 100.0%  learning_rate 3.000e-06  length/mean 282.9  reward/mean 0.6953  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:35:19 left
step 175/300  loss -0.006285  actor/pg_clipfrac 0.001018  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.149e-04  kl 0.04193  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.007962  data_wait_ms 0.02940  grad_norm 0.9144  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,840  samples_per_sec 7.601  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.47  accepted 100.0%  learning_rate 3.000e-06  length/mean 268.7  reward/mean 0.6406  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:35:02 left
step 176/300  loss 0.003929  actor/pg_clipfrac 9.028e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.872e-04  kl 0.03967  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.002342  data_wait_ms 0.03331  grad_norm 0.9400  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,853  samples_per_sec 7.595  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 294.2  reward/mean 0.5391  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:34:45 left
step 177/300  loss 0.01874  actor/pg_clipfrac 5.393e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.329e-04  kl 0.03636  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01729  data_wait_ms 0.02943  grad_norm 0.6658  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,893  samples_per_sec 7.577  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.69  accepted 100.0%  learning_rate 3.000e-06  length/mean 318.7  reward/mean 0.4844  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:34:28 left
step 178/300  loss -0.006645  actor/pg_clipfrac 9.454e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.206e-04  kl 0.03385  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.007999  data_wait_ms 0.03040  grad_norm 0.8274  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,630  samples_per_sec 7.697  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.73  accepted 100.0%  learning_rate 3.000e-06  length/mean 322.3  reward/mean 0.4922  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:34:11 left
step 179/300  loss 0.01673  actor/pg_clipfrac 0.001461  actor/pg_clipfrac_lower 2.756e-05  actor/ppo_kl 3.895e-04  kl 0.03423  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01536  data_wait_ms 0.03153  grad_norm 1.092  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,964  samples_per_sec 7.545  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.46  accepted 100.0%  learning_rate 3.000e-06  length/mean 283.4  reward/mean 0.6484  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:33:54 left
step 180/300  loss 0.02430  actor/pg_clipfrac 6.791e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.604e-04  kl 0.03516  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02290  data_wait_ms 0.02711  grad_norm 0.8916  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,896  samples_per_sec 7.576  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.80  accepted 100.0%  learning_rate 3.000e-06  length/mean 322.1  reward/mean 0.5000  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:33:38 left
step 181/300  loss 0.04164  actor/pg_clipfrac 0.001260  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.799e-04  kl 0.03780  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04013  data_wait_ms 0.02586  grad_norm 0.9527  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,625  samples_per_sec 7.699  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.73  accepted 100.0%  learning_rate 3.000e-06  length/mean 322.5  reward/mean 0.5391  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:33:22 left
step 182/300  loss 0.01986  actor/pg_clipfrac 6.536e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.663e-04  kl 0.03180  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01858  data_wait_ms 0.02961  grad_norm 0.8284  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,957  samples_per_sec 7.548  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.46  accepted 100.0%  learning_rate 3.000e-06  length/mean 286.9  reward/mean 0.6562  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:33:03 left
step 183/300  loss 0.03638  actor/pg_clipfrac 0.001247  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.573e-04  kl 0.03420  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03501  data_wait_ms 0.03155  grad_norm 0.9582  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,615  samples_per_sec 7.704  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.79  accepted 100.0%  learning_rate 3.000e-06  length/mean 338.2  reward/mean 0.6094  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:32:47 left
step 184/300  loss 0.02437  actor/pg_clipfrac 0.001333  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.544e-04  kl 0.03639  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02292  data_wait_ms 0.02970  grad_norm 1.143  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,759  samples_per_sec 7.638  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.45  accepted 100.0%  learning_rate 3.000e-06  length/mean 269.6  reward/mean 0.7344  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:32:29 left
step 185/300  loss 0.01651  actor/pg_clipfrac 0.001030  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.106e-04  kl 0.03172  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01525  data_wait_ms 0.02943  grad_norm 0.9531  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,869  samples_per_sec 7.588  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.59  accepted 100.0%  learning_rate 3.000e-06  length/mean 310.9  reward/mean 0.6641  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:32:11 left
step 186/300  loss 0.01830  actor/pg_clipfrac 0.001192  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.753e-04  kl 0.03734  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01681  data_wait_ms 0.02775  grad_norm 0.9923  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,841  samples_per_sec 7.600  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 308.1  reward/mean 0.6562  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:31:55 left
step 187/300  loss 0.01822  actor/pg_clipfrac 8.837e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.263e-04  kl 0.03847  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01668  data_wait_ms 0.02762  grad_norm 0.7775  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,820  samples_per_sec 7.610  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 327.1  reward/mean 0.4844  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:31:38 left
step 188/300  loss 0.02860  actor/pg_clipfrac 0.001196  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.339e-04  kl 0.03974  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02701  data_wait_ms 0.02888  grad_norm 0.8860  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,871  samples_per_sec 7.587  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 307.0  reward/mean 0.5938  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:31:21 left
step 189/300  loss 0.01484  actor/pg_clipfrac 0.001117  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.473e-04  kl 0.03817  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01332  data_wait_ms 0.03135  grad_norm 0.9911  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,826  samples_per_sec 7.607  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 293.9  reward/mean 0.5234  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:31:04 left
step 190/300  loss 1.728e-04  actor/pg_clipfrac 5.211e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.298e-04  kl 0.03494  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.001225  data_wait_ms 0.02987  grad_norm 0.5432  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,884  samples_per_sec 7.581  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.66  accepted 100.0%  learning_rate 3.000e-06  length/mean 314.9  reward/mean 0.5469  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:30:48 left
step 191/300  loss 0.03598  actor/pg_clipfrac 0.001155  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.580e-04  kl 0.03719  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03449  data_wait_ms 0.02935  grad_norm 0.9042  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,546  samples_per_sec 7.736  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 324.6  reward/mean 0.6484  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:30:31 left
step 192/300  loss 0.03075  actor/pg_clipfrac 0.001211  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.486e-05  kl 0.03720  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02927  data_wait_ms 0.02553  grad_norm 1.027  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,937  samples_per_sec 7.557  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.38  accepted 100.0%  learning_rate 3.000e-06  length/mean 264.5  reward/mean 0.7188  status/truncated 0.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:30:13 left
step 193/300  loss 0.03003  actor/pg_clipfrac 7.119e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.606e-04  kl 0.03601  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02859  data_wait_ms 0.02993  grad_norm 0.8029  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,928  samples_per_sec 7.562  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.77  accepted 100.0%  learning_rate 3.000e-06  length/mean 285.3  reward/mean 0.7031  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:29:58 left
step 194/300  loss -0.001261  actor/pg_clipfrac 0.001214  actor/pg_clipfrac_lower 0  actor/ppo_kl 7.199e-04  kl 0.03496  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.002659  data_wait_ms 0.03008  grad_norm 0.8171  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,610  samples_per_sec 7.706  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.76  accepted 100.0%  learning_rate 3.000e-06  length/mean 321.6  reward/mean 0.5156  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:29:42 left
step 195/300  loss 0.002375  actor/pg_clipfrac 9.791e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.138e-04  kl 0.03640  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 9.190e-04  data_wait_ms 0.02948  grad_norm 0.9764  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,826  samples_per_sec 7.607  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.44  accepted 100.0%  learning_rate 3.000e-06  length/mean 255.3  reward/mean 0.7188  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:29:24 left
step 196/300  loss -0.003309  actor/pg_clipfrac 5.934e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.724e-04  kl 0.04100  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.004949  data_wait_ms 0.02890  grad_norm 0.7670  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,866  samples_per_sec 7.589  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 302.8  reward/mean 0.5078  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:29:07 left
step 197/300  loss 0.02256  actor/pg_clipfrac 4.975e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.880e-04  kl 0.03953  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02097  data_wait_ms 0.02769  grad_norm 0.6850  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,906  samples_per_sec 7.571  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 314.1  reward/mean 0.6016  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:28:51 left
step 198/300  loss 0.01484  actor/pg_clipfrac 7.366e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.089e-04  kl 0.03983  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01325  data_wait_ms 0.02812  grad_norm 0.7926  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,911  samples_per_sec 7.569  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.74  accepted 100.0%  learning_rate 3.000e-06  length/mean 318.2  reward/mean 0.4766  status/truncated 7.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:28:35 left
step 199/300  loss 0.009600  actor/pg_clipfrac 0.001186  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.824e-04  kl 0.03772  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.008091  data_wait_ms 0.02685  grad_norm 0.9575  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,823  samples_per_sec 7.609  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.74  accepted 100.0%  learning_rate 3.000e-06  length/mean 322.7  reward/mean 0.5391  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:28:19 left
step 200/300  loss 0.02719  actor/pg_clipfrac 9.452e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.866e-04  kl 0.03649  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02573  data_wait_ms 0.03054  grad_norm 0.8606  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,781  samples_per_sec 7.628  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 314.1  reward/mean 0.5312  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:28:02 left
step 201/300  loss 0.02883  actor/pg_clipfrac 0.001115  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.607e-04  kl 0.03619  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02738  data_wait_ms 0.02949  grad_norm 0.8346  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,848  samples_per_sec 7.597  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.61  accepted 100.0%  learning_rate 3.000e-06  length/mean 308.2  reward/mean 0.7266  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:27:44 left
step 202/300  loss 0.01362  actor/pg_clipfrac 0.001363  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.846e-04  kl 0.03608  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01218  data_wait_ms 0.02973  grad_norm 0.9618  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,878  samples_per_sec 7.584  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.68  accepted 100.0%  learning_rate 3.000e-06  length/mean 332.5  reward/mean 0.4609  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:27:29 left
step 203/300  loss 0.01764  actor/pg_clipfrac 9.323e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.856e-04  kl 0.03678  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01617  data_wait_ms 0.02789  grad_norm 0.8829  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,810  samples_per_sec 7.615  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 326.8  reward/mean 0.4531  status/truncated 7.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:27:11 left
step 204/300  loss 0.02153  actor/pg_clipfrac 0.001082  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.831e-04  kl 0.03821  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02000  data_wait_ms 0.03021  grad_norm 0.9391  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,732  samples_per_sec 7.650  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 296.1  reward/mean 0.5312  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:26:54 left
step 205/300  loss 0.01094  actor/pg_clipfrac 0.001013  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.275e-04  kl 0.04099  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.009299  data_wait_ms 0.03287  grad_norm 0.8676  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,732  samples_per_sec 7.650  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.56  accepted 100.0%  learning_rate 3.000e-06  length/mean 285.3  reward/mean 0.5859  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:26:38 left
step 206/300  loss 0.03325  actor/pg_clipfrac 0.001232  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.679e-04  kl 0.03534  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03184  data_wait_ms 0.02892  grad_norm 1.021  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,803  samples_per_sec 7.618  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.56  accepted 100.0%  learning_rate 3.000e-06  length/mean 279.0  reward/mean 0.7812  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:26:20 left
step 207/300  loss 0.04583  actor/pg_clipfrac 7.932e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.407e-04  kl 0.03805  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04431  data_wait_ms 0.02724  grad_norm 0.9171  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,793  samples_per_sec 7.622  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 315.2  reward/mean 0.6562  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:26:04 left
step 208/300  loss -0.009587  actor/pg_clipfrac 0.001156  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.074e-04  kl 0.03908  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01115  data_wait_ms 0.02913  grad_norm 0.9150  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,735  samples_per_sec 7.649  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.62  accepted 100.0%  learning_rate 3.000e-06  length/mean 297.5  reward/mean 0.6328  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:25:47 left
step 209/300  loss 0.03574  actor/pg_clipfrac 8.107e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.523e-04  kl 0.04250  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03404  data_wait_ms 0.03022  grad_norm 0.8533  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,757  samples_per_sec 7.638  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.57  accepted 100.0%  learning_rate 3.000e-06  length/mean 308.4  reward/mean 0.4375  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:25:30 left
step 210/300  loss 0.01462  actor/pg_clipfrac 7.847e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.858e-04  kl 0.03510  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01322  data_wait_ms 0.03002  grad_norm 0.8402  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,915  samples_per_sec 7.567  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.59  accepted 100.0%  learning_rate 3.000e-06  length/mean 298.7  reward/mean 0.5391  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:25:13 left
step 211/300  loss 0.02728  actor/pg_clipfrac 6.241e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.385e-04  kl 0.03716  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02579  data_wait_ms 0.03113  grad_norm 0.9320  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,848  samples_per_sec 7.597  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.75  accepted 100.0%  learning_rate 3.000e-06  length/mean 275.4  reward/mean 0.6641  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:24:56 left
step 212/300  loss -0.002526  actor/pg_clipfrac 8.420e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.208e-04  kl 0.03692  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.004003  data_wait_ms 0.02700  grad_norm 0.8098  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,891  samples_per_sec 7.578  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.68  accepted 100.0%  learning_rate 3.000e-06  length/mean 315.5  reward/mean 0.6016  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:24:40 left
step 213/300  loss 0.03586  actor/pg_clipfrac 0.001477  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.193e-04  kl 0.03721  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03437  data_wait_ms 0.03042  grad_norm 0.9784  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,784  samples_per_sec 7.626  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 312.1  reward/mean 0.5781  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:24:23 left
step 214/300  loss -0.01070  actor/pg_clipfrac 0.001314  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.729e-04  kl 0.04107  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01234  data_wait_ms 0.02961  grad_norm 1.058  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,770  samples_per_sec 7.633  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.61  accepted 100.0%  learning_rate 3.000e-06  length/mean 267.6  reward/mean 0.5234  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:24:06 left
step 215/300  loss 0.01160  actor/pg_clipfrac 6.767e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.339e-04  kl 0.03691  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01013  data_wait_ms 0.02951  grad_norm 0.7188  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,745  samples_per_sec 7.644  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.60  accepted 100.0%  learning_rate 3.000e-06  length/mean 288.6  reward/mean 0.6562  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:23:50 left
step 216/300  loss 0.02602  actor/pg_clipfrac 6.901e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.968e-04  kl 0.04358  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02427  data_wait_ms 0.02938  grad_norm 0.7897  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,772  samples_per_sec 7.632  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.58  accepted 100.0%  learning_rate 3.000e-06  length/mean 283.0  reward/mean 0.4844  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:23:32 left
step 217/300  loss 0.04887  actor/pg_clipfrac 9.109e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 7.269e-04  kl 0.04301  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04715  data_wait_ms 0.03240  grad_norm 0.9000  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,680  samples_per_sec 7.674  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.60  accepted 100.0%  learning_rate 3.000e-06  length/mean 300.2  reward/mean 0.6328  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:23:15 left
step 218/300  loss 0.03072  actor/pg_clipfrac 0.001140  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.937e-04  kl 0.03928  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02915  data_wait_ms 0.02951  grad_norm 0.8667  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,882  samples_per_sec 7.582  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.51  accepted 100.0%  learning_rate 3.000e-06  length/mean 267.2  reward/mean 0.7031  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:22:58 left
step 219/300  loss 0.04201  actor/pg_clipfrac 0.001208  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.692e-04  kl 0.04298  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04029  data_wait_ms 0.02975  grad_norm 0.9725  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,858  samples_per_sec 7.593  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 297.6  reward/mean 0.5156  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:22:42 left
step 220/300  loss 0.05102  actor/pg_clipfrac 0.001313  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.284e-04  kl 0.04497  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04923  data_wait_ms 0.03570  grad_norm 0.9904  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,747  samples_per_sec 7.643  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.69  accepted 100.0%  learning_rate 3.000e-06  length/mean 309.3  reward/mean 0.5312  status/truncated 7.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:22:25 left
step 221/300  loss 0.01467  actor/pg_clipfrac 7.630e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.232e-04  kl 0.03969  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01309  data_wait_ms 0.02877  grad_norm 0.9373  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,927  samples_per_sec 7.562  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.58  accepted 100.0%  learning_rate 3.000e-06  length/mean 276.4  reward/mean 0.5156  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:22:08 left
step 222/300  loss 0.03754  actor/pg_clipfrac 0.001123  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.973e-04  kl 0.03882  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03598  data_wait_ms 0.02784  grad_norm 0.8985  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,726  samples_per_sec 7.653  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.76  accepted 100.0%  learning_rate 3.000e-06  length/mean 306.0  reward/mean 0.5781  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:21:51 left
step 223/300  loss -0.008749  actor/pg_clipfrac 0.001034  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.894e-04  kl 0.04382  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01050  data_wait_ms 0.02689  grad_norm 0.8929  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,868  samples_per_sec 7.588  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.56  accepted 100.0%  learning_rate 3.000e-06  length/mean 272.0  reward/mean 0.6172  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:21:34 left
step 224/300  loss 0.04724  actor/pg_clipfrac 0.001189  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.471e-04  kl 0.04210  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04556  data_wait_ms 0.03126  grad_norm 0.9079  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,872  samples_per_sec 7.587  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 289.1  reward/mean 0.5781  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:21:17 left
step 225/300  loss 0.002257  actor/pg_clipfrac 6.193e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.644e-04  kl 0.04765  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 3.506e-04  data_wait_ms 0.03081  grad_norm 0.7854  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,901  samples_per_sec 7.573  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 315.4  reward/mean 0.4297  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:21:00 left
step 226/300  loss -0.01170  actor/pg_clipfrac 5.482e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.389e-04  kl 0.04255  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01340  data_wait_ms 0.02986  grad_norm 0.6339  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,889  samples_per_sec 7.579  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.73  accepted 100.0%  learning_rate 3.000e-06  length/mean 313.5  reward/mean 0.4766  status/truncated 10.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:20:44 left
step 227/300  loss 0.01532  actor/pg_clipfrac 4.783e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.262e-04  kl 0.04428  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01355  data_wait_ms 0.03147  grad_norm 0.7765  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,608  samples_per_sec 7.707  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.72  accepted 100.0%  learning_rate 3.000e-06  length/mean 277.7  reward/mean 0.6953  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:20:27 left
step 228/300  loss -8.829e-04  actor/pg_clipfrac 5.471e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.439e-04  kl 0.04191  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.002559  data_wait_ms 0.02925  grad_norm 0.7687  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,857  samples_per_sec 7.593  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.44  accepted 100.0%  learning_rate 3.000e-06  length/mean 271.3  reward/mean 0.6797  status/truncated 0.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:20:10 left
step 229/300  loss 0.03237  actor/pg_clipfrac 7.502e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.481e-04  kl 0.04273  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03066  data_wait_ms 0.02776  grad_norm 0.8501  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,674  samples_per_sec 7.677  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.69  accepted 100.0%  learning_rate 3.000e-06  length/mean 302.0  reward/mean 0.4375  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:19:54 left
step 230/300  loss 0.02302  actor/pg_clipfrac 7.164e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 7.780e-04  kl 0.04830  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02109  data_wait_ms 0.03071  grad_norm 0.9170  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,874  samples_per_sec 7.585  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.50  accepted 100.0%  learning_rate 3.000e-06  length/mean 283.5  reward/mean 0.5078  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:19:36 left
step 231/300  loss 0.01132  actor/pg_clipfrac 5.574e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.769e-04  kl 0.04656  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.009457  data_wait_ms 0.02692  grad_norm 0.8515  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,874  samples_per_sec 7.586  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 294.3  reward/mean 0.6172  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:19:20 left
step 232/300  loss 0.03155  actor/pg_clipfrac 0.001062  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.157e-04  kl 0.04348  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02981  data_wait_ms 0.02984  grad_norm 0.8196  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,816  samples_per_sec 7.612  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 330.9  reward/mean 0.4688  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:19:03 left
step 233/300  loss 0.003635  actor/pg_clipfrac 4.372e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 7.475e-04  kl 0.04360  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.001891  data_wait_ms 0.02656  grad_norm 0.7789  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,834  samples_per_sec 7.604  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 303.8  reward/mean 0.3828  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:18:47 left
step 234/300  loss 0.03351  actor/pg_clipfrac 0.001029  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.464e-04  kl 0.04675  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03164  data_wait_ms 0.02765  grad_norm 0.8528  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,803  samples_per_sec 7.618  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 311.4  reward/mean 0.4922  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:18:30 left
step 235/300  loss 0.009733  actor/pg_clipfrac 8.111e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.699e-04  kl 0.04718  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.007846  data_wait_ms 0.02753  grad_norm 0.8263  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,724  samples_per_sec 7.654  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 308.2  reward/mean 0.5234  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:18:13 left
step 236/300  loss 0.03005  actor/pg_clipfrac 6.956e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.382e-04  kl 0.04640  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02819  data_wait_ms 0.02997  grad_norm 0.8625  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,793  samples_per_sec 7.622  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.56  accepted 100.0%  learning_rate 3.000e-06  length/mean 269.6  reward/mean 0.7500  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:17:55 left
step 237/300  loss 0.02185  actor/pg_clipfrac 7.087e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.697e-04  kl 0.04371  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02010  data_wait_ms 0.03070  grad_norm 0.8877  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,856  samples_per_sec 7.594  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.63  accepted 100.0%  learning_rate 3.000e-06  length/mean 297.6  reward/mean 0.6406  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:17:39 left
step 238/300  loss -0.007730  actor/pg_clipfrac 6.726e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.837e-04  kl 0.04638  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.009585  data_wait_ms 0.02820  grad_norm 0.8289  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,765  samples_per_sec 7.635  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.69  accepted 100.0%  learning_rate 3.000e-06  length/mean 290.4  reward/mean 0.4609  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:17:22 left
step 239/300  loss 0.01797  actor/pg_clipfrac 7.386e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.188e-04  kl 0.04705  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01609  data_wait_ms 0.03400  grad_norm 0.9273  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,901  samples_per_sec 7.574  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.60  accepted 100.0%  learning_rate 3.000e-06  length/mean 285.6  reward/mean 0.6797  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:17:05 left
step 240/300  loss 0.02581  actor/pg_clipfrac 0.001338  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.109e-04  kl 0.04430  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02404  data_wait_ms 0.03386  grad_norm 0.9779  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,907  samples_per_sec 7.571  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.73  accepted 100.0%  learning_rate 3.000e-06  length/mean 332.8  reward/mean 0.4844  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:16:49 left
step 241/300  loss 0.01630  actor/pg_clipfrac 5.678e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.031e-04  kl 0.04258  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01460  data_wait_ms 0.02816  grad_norm 0.9668  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,819  samples_per_sec 7.611  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.74  accepted 100.0%  learning_rate 3.000e-06  length/mean 275.2  reward/mean 0.6250  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:16:33 left
step 242/300  loss 0.007637  actor/pg_clipfrac 7.050e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.778e-04  kl 0.04427  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.005866  data_wait_ms 0.02962  grad_norm 0.7617  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,757  samples_per_sec 7.638  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 310.3  reward/mean 0.5391  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:16:16 left
step 243/300  loss -0.001168  actor/pg_clipfrac 2.588e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.856e-04  kl 0.04478  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.002960  data_wait_ms 0.02868  grad_norm 0.5244  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,861  samples_per_sec 7.592  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.59  accepted 100.0%  learning_rate 3.000e-06  length/mean 271.7  reward/mean 0.6484  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:15:59 left
step 244/300  loss 0.03661  actor/pg_clipfrac 8.535e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.947e-04  kl 0.04474  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03482  data_wait_ms 0.02867  grad_norm 0.8376  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,759  samples_per_sec 7.638  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.69  accepted 100.0%  learning_rate 3.000e-06  length/mean 320.4  reward/mean 0.4922  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:15:42 left
step 245/300  loss 0.004533  actor/pg_clipfrac 0.001022  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.572e-04  kl 0.04579  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.002702  data_wait_ms 0.03075  grad_norm 0.8850  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,936  samples_per_sec 7.558  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.59  accepted 100.0%  learning_rate 3.000e-06  length/mean 305.9  reward/mean 0.5859  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:15:25 left
step 246/300  loss 0.001460  actor/pg_clipfrac 0.001070  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.517e-04  kl 0.04868  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -4.869e-04  data_wait_ms 0.03021  grad_norm 0.9085  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,722  samples_per_sec 7.655  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.77  accepted 100.0%  learning_rate 3.000e-06  length/mean 335.8  reward/mean 0.4766  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:15:08 left
step 247/300  loss 0.01764  actor/pg_clipfrac 8.581e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.629e-04  kl 0.04461  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01586  data_wait_ms 0.02766  grad_norm 0.9083  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,772  samples_per_sec 7.632  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.55  accepted 100.0%  learning_rate 3.000e-06  length/mean 282.2  reward/mean 0.7109  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:14:51 left
step 248/300  loss 0.02585  actor/pg_clipfrac 7.069e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.603e-04  kl 0.04282  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02414  data_wait_ms 0.03023  grad_norm 0.7599  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,794  samples_per_sec 7.622  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.60  accepted 100.0%  learning_rate 3.000e-06  length/mean 287.3  reward/mean 0.6250  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:14:34 left
step 249/300  loss 0.01383  actor/pg_clipfrac 0.001533  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.284e-04  kl 0.04712  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01195  data_wait_ms 0.02951  grad_norm 0.9977  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,834  samples_per_sec 7.604  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.63  accepted 100.0%  learning_rate 3.000e-06  length/mean 295.6  reward/mean 0.6172  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:14:17 left
step 250/300  loss 0.01619  actor/pg_clipfrac 8.748e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.515e-04  kl 0.04581  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01436  data_wait_ms 0.03448  grad_norm 0.9082  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,608  samples_per_sec 7.707  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.66  accepted 100.0%  learning_rate 3.000e-06  length/mean 294.7  reward/mean 0.6641  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:14:00 left
step 251/300  loss -0.002707  actor/pg_clipfrac 0.001182  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.653e-04  kl 0.04659  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.004571  data_wait_ms 0.02974  grad_norm 0.9805  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,783  samples_per_sec 7.627  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.44  accepted 100.0%  learning_rate 3.000e-06  length/mean 277.7  reward/mean 0.5703  status/truncated 0.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:13:43 left
step 252/300  loss 0.02007  actor/pg_clipfrac 9.887e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.765e-04  kl 0.05079  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01804  data_wait_ms 0.03118  grad_norm 0.8986  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,713  samples_per_sec 7.659  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.61  accepted 100.0%  learning_rate 3.000e-06  length/mean 284.5  reward/mean 0.5625  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:13:27 left
step 253/300  loss 0.005293  actor/pg_clipfrac 7.860e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.918e-06  kl 0.05153  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.003232  data_wait_ms 0.02717  grad_norm 0.8897  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,796  samples_per_sec 7.621  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.55  accepted 100.0%  learning_rate 3.000e-06  length/mean 298.2  reward/mean 0.4688  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:13:10 left
step 254/300  loss 0.01214  actor/pg_clipfrac 0.001263  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.917e-04  kl 0.05021  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01014  data_wait_ms 0.03049  grad_norm 0.9161  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,917  samples_per_sec 7.566  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.63  accepted 100.0%  learning_rate 3.000e-06  length/mean 290.8  reward/mean 0.5859  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:12:53 left
step 255/300  loss 0.003307  actor/pg_clipfrac 0.001293  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.122e-04  kl 0.04699  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.001428  data_wait_ms 0.02724  grad_norm 0.9407  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,994  samples_per_sec 7.532  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.75  accepted 100.0%  learning_rate 3.000e-06  length/mean 302.1  reward/mean 0.5781  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:12:36 left
step 256/300  loss 0.03978  actor/pg_clipfrac 5.120e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.309e-04  kl 0.04344  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03804  data_wait_ms 0.02620  grad_norm 0.7447  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,852  samples_per_sec 7.596  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.82  accepted 100.0%  learning_rate 3.000e-06  length/mean 335.7  reward/mean 0.5234  status/truncated 17.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:12:20 left
step 257/300  loss 0.04558  actor/pg_clipfrac 0.001354  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.261e-04  kl 0.04928  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04361  data_wait_ms 0.03410  grad_norm 0.9626  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,861  samples_per_sec 7.591  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.69  accepted 100.0%  learning_rate 3.000e-06  length/mean 328.9  reward/mean 0.4609  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:12:03 left
step 258/300  loss 0.03676  actor/pg_clipfrac 8.365e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.243e-04  kl 0.04786  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03485  data_wait_ms 0.03256  grad_norm 0.7045  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,800  samples_per_sec 7.619  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.69  accepted 100.0%  learning_rate 3.000e-06  length/mean 336.2  reward/mean 0.4219  status/truncated 8.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:11:46 left
step 259/300  loss 0.02750  actor/pg_clipfrac 8.597e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.633e-04  kl 0.04641  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02564  data_wait_ms 0.02647  grad_norm 0.9208  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,929  samples_per_sec 7.561  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.63  accepted 100.0%  learning_rate 3.000e-06  length/mean 290.8  reward/mean 0.6641  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:11:30 left
step 260/300  loss 0.02310  actor/pg_clipfrac 6.610e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.658e-04  kl 0.05068  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02108  data_wait_ms 0.02944  grad_norm 0.7261  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,896  samples_per_sec 7.576  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.76  accepted 100.0%  learning_rate 3.000e-06  length/mean 331.0  reward/mean 0.5078  status/truncated 11.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:11:13 left
step 261/300  loss 0.02957  actor/pg_clipfrac 0.001028  actor/pg_clipfrac_lower 0  actor/ppo_kl 9.239e-05  kl 0.04505  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02777  data_wait_ms 0.02934  grad_norm 0.9266  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,798  samples_per_sec 7.620  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.73  accepted 100.0%  learning_rate 3.000e-06  length/mean 311.5  reward/mean 0.6250  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:10:56 left
step 262/300  loss 0.03095  actor/pg_clipfrac 8.892e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.942e-04  kl 0.04790  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02903  data_wait_ms 0.03002  grad_norm 0.7314  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,912  samples_per_sec 7.568  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.63  accepted 100.0%  learning_rate 3.000e-06  length/mean 316.3  reward/mean 0.5859  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:10:39 left
step 263/300  loss 0.04529  actor/pg_clipfrac 0.001035  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.148e-04  kl 0.04975  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04330  data_wait_ms 0.02533  grad_norm 0.9525  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,923  samples_per_sec 7.564  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.74  accepted 100.0%  learning_rate 3.000e-06  length/mean 309.5  reward/mean 0.6641  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:10:23 left
step 264/300  loss 0.04211  actor/pg_clipfrac 8.600e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.123e-04  kl 0.04968  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04013  data_wait_ms 0.02809  grad_norm 1.001  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 17,015  samples_per_sec 7.523  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.75  accepted 100.0%  learning_rate 3.000e-06  length/mean 308.9  reward/mean 0.5469  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:10:06 left
step 265/300  loss -0.003811  actor/pg_clipfrac 5.976e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.808e-04  kl 0.04638  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.005666  data_wait_ms 0.02867  grad_norm 0.6344  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,988  samples_per_sec 7.535  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.85  accepted 100.0%  learning_rate 3.000e-06  length/mean 313.8  reward/mean 0.4688  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:09:49 left
step 266/300  loss 0.04137  actor/pg_clipfrac 8.480e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.698e-04  kl 0.04509  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03956  data_wait_ms 0.02990  grad_norm 0.7880  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,757  samples_per_sec 7.639  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.82  accepted 100.0%  learning_rate 3.000e-06  length/mean 340.9  reward/mean 0.5859  status/truncated 13.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:09:33 left
step 267/300  loss 0.008438  actor/pg_clipfrac 7.398e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.513e-04  kl 0.04503  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.006637  data_wait_ms 0.03092  grad_norm 0.8754  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,919  samples_per_sec 7.566  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.59  accepted 100.0%  learning_rate 3.000e-06  length/mean 295.7  reward/mean 0.6094  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:09:16 left
step 268/300  loss 0.03149  actor/pg_clipfrac 8.835e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.507e-04  kl 0.04560  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02967  data_wait_ms 0.04502  grad_norm 0.7928  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,928  samples_per_sec 7.561  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.75  accepted 100.0%  learning_rate 3.000e-06  length/mean 362.5  reward/mean 0.3594  status/truncated 10.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:08:59 left
step 269/300  loss 0.02504  actor/pg_clipfrac 0.001046  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.192e-04  kl 0.04473  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02325  data_wait_ms 0.03481  grad_norm 0.9053  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,854  samples_per_sec 7.595  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.76  accepted 100.0%  learning_rate 3.000e-06  length/mean 343.5  reward/mean 0.4375  status/truncated 10.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:08:42 left
step 270/300  loss 0.02169  actor/pg_clipfrac 0.001081  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.731e-04  kl 0.04479  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01990  data_wait_ms 0.03588  grad_norm 0.9046  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,870  samples_per_sec 7.588  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.68  accepted 100.0%  learning_rate 3.000e-06  length/mean 310.8  reward/mean 0.5703  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:08:26 left
step 271/300  loss -0.008559  actor/pg_clipfrac 6.844e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 7.384e-04  kl 0.04293  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01028  data_wait_ms 0.03008  grad_norm 0.9048  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,973  samples_per_sec 7.541  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 331.0  reward/mean 0.5078  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:08:09 left
step 272/300  loss 0.02206  actor/pg_clipfrac 9.703e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.746e-04  kl 0.04108  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02041  data_wait_ms 0.03723  grad_norm 0.7527  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,915  samples_per_sec 7.567  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.80  accepted 100.0%  learning_rate 3.000e-06  length/mean 330.1  reward/mean 0.5312  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:07:52 left
step 273/300  loss 0.009767  actor/pg_clipfrac 9.894e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.154e-04  kl 0.04412  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.008002  data_wait_ms 0.02927  grad_norm 0.8547  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,907  samples_per_sec 7.571  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.75  accepted 100.0%  learning_rate 3.000e-06  length/mean 323.8  reward/mean 0.4844  status/truncated 10.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:07:35 left
step 274/300  loss 0.02037  actor/pg_clipfrac 5.607e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.550e-04  kl 0.04197  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01870  data_wait_ms 0.03353  grad_norm 0.6271  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,866  samples_per_sec 7.589  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.74  accepted 100.0%  learning_rate 3.000e-06  length/mean 334.4  reward/mean 0.5312  status/truncated 10.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:07:19 left
step 275/300  loss 0.03975  actor/pg_clipfrac 0.001288  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.003e-04  kl 0.04744  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03785  data_wait_ms 0.02917  grad_norm 0.9457  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 17,019  samples_per_sec 7.521  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 339.8  reward/mean 0.4531  status/truncated 7.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:07:02 left
step 276/300  loss 0.03859  actor/pg_clipfrac 7.452e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.965e-04  kl 0.04652  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03673  data_wait_ms 0.02625  grad_norm 0.7370  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,820  samples_per_sec 7.610  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.85  accepted 100.0%  learning_rate 3.000e-06  length/mean 314.5  reward/mean 0.4141  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:06:46 left
step 277/300  loss 0.01991  actor/pg_clipfrac 7.817e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.381e-04  kl 0.04381  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01816  data_wait_ms 0.02908  grad_norm 0.8760  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,802  samples_per_sec 7.618  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 299.8  reward/mean 0.5234  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:06:29 left
step 278/300  loss 0.01224  actor/pg_clipfrac 9.657e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.131e-04  kl 0.04558  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01041  data_wait_ms 0.02984  grad_norm 0.8616  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,866  samples_per_sec 7.589  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.63  accepted 100.0%  learning_rate 3.000e-06  length/mean 299.3  reward/mean 0.5625  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:06:12 left
step 279/300  loss 0.02683  actor/pg_clipfrac 9.828e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.957e-04  kl 0.04583  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02499  data_wait_ms 0.02643  grad_norm 0.8752  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,814  samples_per_sec 7.613  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.70  accepted 100.0%  learning_rate 3.000e-06  length/mean 294.1  reward/mean 0.5938  status/truncated 6.2%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:05:55 left
step 280/300  loss 0.02615  actor/pg_clipfrac 9.154e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.014e-04  kl 0.04918  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02418  data_wait_ms 0.03249  grad_norm 0.8935  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,874  samples_per_sec 7.586  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 307.2  reward/mean 0.5781  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:05:38 left
step 281/300  loss 0.01556  actor/pg_clipfrac 7.666e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.301e-04  kl 0.03910  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01399  data_wait_ms 0.02986  grad_norm 0.8420  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,769  samples_per_sec 7.633  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 275.2  reward/mean 0.7266  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:05:21 left
step 282/300  loss 0.007004  actor/pg_clipfrac 9.784e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.742e-04  kl 0.04401  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.005244  data_wait_ms 0.03279  grad_norm 0.8686  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,837  samples_per_sec 7.602  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.60  accepted 100.0%  learning_rate 3.000e-06  length/mean 303.4  reward/mean 0.7031  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:05:04 left
step 283/300  loss -0.02343  actor/pg_clipfrac 0.001015  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.279e-04  kl 0.04453  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.02521  data_wait_ms 0.02998  grad_norm 0.8660  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,928  samples_per_sec 7.562  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.67  accepted 100.0%  learning_rate 3.000e-06  length/mean 323.3  reward/mean 0.4766  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:04:47 left
step 284/300  loss 0.01851  actor/pg_clipfrac 8.993e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.992e-04  kl 0.04074  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01688  data_wait_ms 0.02997  grad_norm 0.7262  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,879  samples_per_sec 7.583  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.76  accepted 100.0%  learning_rate 3.000e-06  length/mean 330.1  reward/mean 0.4922  status/truncated 12.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:04:30 left
step 285/300  loss 0.003646  actor/pg_clipfrac 9.945e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.969e-04  kl 0.04285  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.001932  data_wait_ms 0.02891  grad_norm 0.9182  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,694  samples_per_sec 7.667  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.71  accepted 100.0%  learning_rate 3.000e-06  length/mean 314.2  reward/mean 0.7422  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:04:13 left
step 286/300  loss 0.04763  actor/pg_clipfrac 0.001399  actor/pg_clipfrac_lower 0  actor/ppo_kl 8.159e-04  kl 0.04543  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.04581  data_wait_ms 0.02873  grad_norm 1.092  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,814  samples_per_sec 7.613  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.53  accepted 100.0%  learning_rate 3.000e-06  length/mean 296.0  reward/mean 0.5938  status/truncated 0.8%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:03:56 left
step 287/300  loss 0.03450  actor/pg_clipfrac 7.572e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.545e-04  kl 0.03952  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03292  data_wait_ms 0.02853  grad_norm 0.8542  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,763  samples_per_sec 7.636  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.65  accepted 100.0%  learning_rate 3.000e-06  length/mean 299.2  reward/mean 0.6641  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:03:39 left
step 288/300  loss -0.003528  actor/pg_clipfrac 0.001000  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.553e-04  kl 0.03999  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.005128  data_wait_ms 0.03024  grad_norm 0.8847  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,810  samples_per_sec 7.615  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.60  accepted 100.0%  learning_rate 3.000e-06  length/mean 304.7  reward/mean 0.6484  status/truncated 4.7%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:03:22 left
step 289/300  loss 0.006119  actor/pg_clipfrac 6.322e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.046e-04  kl 0.04599  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.004279  data_wait_ms 0.04682  grad_norm 0.7632  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,773  samples_per_sec 7.632  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 271.9  reward/mean 0.5469  status/truncated 2.3%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:03:05 left
step 290/300  loss -0.006522  actor/pg_clipfrac 8.877e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.809e-04  kl 0.03809  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.008046  data_wait_ms 0.03238  grad_norm 0.6841  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,936  samples_per_sec 7.558  hlo_flops_util_lower_bound 8.4%  hfu_lower_bound 9.4%  unsized_kernels 291.0  rollout_seconds 14.60  accepted 100.0%  learning_rate 3.000e-06  length/mean 325.6  reward/mean 0.4609  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:02:49 left
step 291/300  loss 0.03358  actor/pg_clipfrac 0.001027  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.639e-04  kl 0.03767  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.03208  data_wait_ms 0.02806  grad_norm 0.7786  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,811  samples_per_sec 7.614  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.77  accepted 100.0%  learning_rate 3.000e-06  length/mean 311.9  reward/mean 0.5312  status/truncated 9.4%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:02:32 left
step 292/300  loss 0.02655  actor/pg_clipfrac 4.798e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.978e-04  kl 0.04072  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02492  data_wait_ms 0.02822  grad_norm 0.6635  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,631  samples_per_sec 7.696  hlo_flops_util_lower_bound 8.6%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.64  accepted 100.0%  learning_rate 3.000e-06  length/mean 276.8  reward/mean 0.7344  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:02:15 left
step 293/300  loss -0.009385  actor/pg_clipfrac 0.001212  actor/pg_clipfrac_lower 0  actor/ppo_kl 7.988e-04  kl 0.04464  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01117  data_wait_ms 0.02994  grad_norm 0.9751  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,659  samples_per_sec 7.683  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.46  accepted 100.0%  learning_rate 3.000e-06  length/mean 270.6  reward/mean 0.5703  status/truncated 0.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:01:58 left
step 294/300  loss 0.01107  actor/pg_clipfrac 0.001237  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.602e-04  kl 0.04400  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.009315  data_wait_ms 0.02578  grad_norm 0.9874  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,797  samples_per_sec 7.620  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.49  accepted 100.0%  learning_rate 3.000e-06  length/mean 271.6  reward/mean 0.6016  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:01:41 left
step 295/300  loss 0.01590  actor/pg_clipfrac 7.978e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 2.852e-04  kl 0.04081  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01427  data_wait_ms 0.02921  grad_norm 0.7754  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,777  samples_per_sec 7.629  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.63  accepted 100.0%  learning_rate 3.000e-06  length/mean 293.8  reward/mean 0.5938  status/truncated 3.1%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:01:24 left
step 296/300  loss 0.02110  actor/pg_clipfrac 5.549e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 5.728e-05  kl 0.04712  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01922  data_wait_ms 0.02645  grad_norm 0.8327  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,659  samples_per_sec 7.683  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.6%  unsized_kernels 291.0  rollout_seconds 14.61  accepted 100.0%  learning_rate 3.000e-06  length/mean 281.6  reward/mean 0.4531  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:01:07 left
step 297/300  loss 0.02480  actor/pg_clipfrac 0.001185  actor/pg_clipfrac_lower 0  actor/ppo_kl 6.694e-04  kl 0.04658  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.02293  data_wait_ms 0.02983  grad_norm 1.060  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,762  samples_per_sec 7.636  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.49  accepted 100.0%  learning_rate 3.000e-06  length/mean 263.6  reward/mean 0.6250  status/truncated 0.0%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:00:50 left
step 298/300  loss 0.01245  actor/pg_clipfrac 5.430e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 1.696e-04  kl 0.04575  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01062  data_wait_ms 0.04541  grad_norm 0.8477  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,705  samples_per_sec 7.662  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.59  accepted 100.0%  learning_rate 3.000e-06  length/mean 316.5  reward/mean 0.6094  status/truncated 5.5%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:00:34 left
step 299/300  loss 0.01311  actor/pg_clipfrac 9.982e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 4.219e-04  kl 0.05370  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg 0.01096  data_wait_ms 0.02749  grad_norm 0.9304  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 16,792  samples_per_sec 7.623  hlo_flops_util_lower_bound 8.5%  hfu_lower_bound 9.5%  unsized_kernels 291.0  rollout_seconds 14.54  accepted 100.0%  learning_rate 3.000e-06  length/mean 273.9  reward/mean 0.5312  status/truncated 1.6%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  0:00:17 left
step 300/300  loss -0.01090  actor/pg_clipfrac 9.158e-04  actor/pg_clipfrac_lower 0  actor/ppo_kl 3.467e-04  kl 0.04809  mismatch/ess 1.000  mismatch/k3_kl 0  mismatch/kl 0  pg -0.01282  data_wait_ms 0.02608  grad_norm 0.9309  compiles 1.000  cache_hits 0  cache_misses 0  peak_memory_gib 24.30  step_time_ms 3,041  samples_per_sec 42.10  hlo_flops_util_lower_bound 46.8%  hfu_lower_bound 52.4%  unsized_kernels 291.0  rollout_seconds 14.62  accepted 100.0%  learning_rate 3.000e-06  length/mean 298.6  reward/mean 0.5781  status/truncated 3.9%  accepted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,472  attempted/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 7,473  chunks_per_record/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0  rejected/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3/over_length 1.000  truncated/memory:6bc361644c6a032d8f5b721f528fdd7ea56e0305448c8749d34237c0effa90e3 0
Trained 300 steps in 1:25:49: first step after 117.86 s, then 0.1 step/s
97.7% of the wall time in steps, final loss -0.01090
share of replies against step00.250.50.751100200300stepshare of repliesmean reward 0.58cut off 0.04

Each point is one batch of 16 questions, so the faint line jumps around a lot; the thick one is a running mean over 25 steps. It climbs from about 0.51 to a peak near 0.63 around step 150, then settles back to about 0.58. The share of replies cut off at 512 tokens rises to about 7% late in training, with single batches as high as 17%, and ends near 4%. But this reward is measured on the questions the model is learning from, sampled at temperature 1. Does the improvement carry over to questions it never trained on?

The model after training

Section titled “The model after training”
after, right_after = grade(state.variables, test)
print(f"right before training {right_before.mean():.1%}, "
      f"after {right_after.mean():.1%}")




def outcomes(replies, right):
    unboxed = np.mean([r"\boxed{" not in reply for reply in replies])
    return [right.mean(), 1 - right.mean() - unboxed, unboxed]




figures.bars([outcomes(before, right_before), outcomes(after, right_after)],
             labels=["right", "wrong", "no answer"],
             captions=["before training", "after training"],
             colors=["accent", "warm", "tint1"], y_label="share of test replies")
right before training 48.7%, after 52.1%
share of test replies: right, wrong, no answershare of test repliesright0.490.52wrong0.460.45no answer0.050.03before trainingafter training
def outcome(reply, right):
    return "right" if right else "wrong" if r"\boxed{" in reply else "no answer"




kinds = ["right", "wrong", "no answer"]
moves = Counter((outcome(*then), outcome(*now)) for then, now in
                zip(zip(before, right_before), zip(after, right_after)))
print(f"{'before, after':>14}" + "".join(f"{kind:>11}" for kind in kinds))
for then in kinds:
    print(f"{then:>14}" + "".join(f"{moves[then, now]:>11}" for now in kinds))
 before, after      right      wrong  no answer
         right        497        139          7
         wrong        174        417         19
     no answer         16         35         15

Right answers go from 48.7% to 52.1%, but the table shows that this is a net change. Of the 643 questions the model got right before training, it now misses 146, and it picks up 190 it used to miss, 16 of them replies that never boxed an answer. Training moves questions both ways, and the bars show the difference.

fixed = next(i for i in range(len(test)) if right_after[i] > right_before[i])
print(test[fixed]["question"], "\nreference:", reference(test[fixed]))
print("\n--- before ---\n" + before[fixed])
print("\n--- after ---\n" + after[fixed])
A robe takes 2 bolts of blue fiber and half that much white fiber.  How many bolts in total does it take? 
reference: 3

--- before ---
To determine the total number of bolts of blue and white fiber needed, we start by defining the variables and setting up the equation based on the given information.

Let:
- \( B \) be the number of bolts of blue fiber.
- \( W \) be the number of bolts of white fiber.

According to the problem, the robe takes 2 bolts of blue fiber and half that much white fiber. Therefore, we can write the following equation:

\[ 2B = W \]

We need to find the total number of bolts, which is the sum of the blue and white fibers:

\[ B + W \]

To find \( B + W \), we can substitute \( W \) from the equation \( 2B = W \) into the total equation:

\[ B + W = B + 2B \]

Simplify the right-hand side:

\[ B + 2B = 3B \]

Thus, the total number of bolts needed is:

\[ \boxed{3B} \]

--- after ---
To determine the total number of bolts needed, we need to break down the problem step by step.

1. **Identify the amount of blue fiber needed:**
   - The problem states that it takes 2 bolts of blue fiber.

2. **Determine the amount of white fiber needed:**
   - The problem states that it takes half as much white fiber as blue fiber.
   - Therefore, the amount of white fiber needed is \( \frac{2}{2} = 1 \) bolt.

3. **Calculate the total number of bolts needed:**
   - To find the total number of bolts, we add the number of blue fiber bolts to the number of white fiber bolts.
   - The total number of bolts is \( 2 + 1 = 3 \).

Thus, the total number of bolts needed is \(\boxed{3}\).

What went wrong before training? The model reversed the relation, writing 2B=W2B = W (twice as much white fiber, where the question says half), and boxed 3B3B, which isn’t even a number. After training it reads the numbers straight off the question: 2 bolts of blue, half of that is 1 bolt of white, 3 in all. All the model ever got back during training was a 1 or a 0 per reply.

What changes at a larger scale

Section titled “What changes at a larger scale”

The loss stays the same. A bigger model gets its logits from a deeper network, but GRPOObjective computes the same per-token terms we just checked. What grows is everything around it. Rewards like running a program’s tests need a sandbox (dew.rl.sandbox). While the trainer updates the model, a separate engine (Dew’s Server, vLLM or SGLang) generates replies, so they can be one update old by the time we train on them. The example scores them again with the weights the update starts from, and caps each token’s importance weight with behavior_importance. And long or multi-turn episodes are packed into rows with masks, which tutorial 11 walks through.

examples/train_rlvr.py puts all of this together and trains a model to write Python programs that pass their tests. Post-training covers Dew’s other objectives.