on-policy-self-distillation

gpt-5.6-sol

on-policy-self-distillation — graded by the task's verifier against a measured baseline.

backend: modal
variant: codex azure_ai/gpt-5.6-sol
metric: AIME24 avg@12 accuracy
actions: 627
duration: 6h 30m
recorded: 2026-08-02

run idopsd-aime24-4gpus__RqGcDty

Everything gpt-5.6-sol did on this task, in order.

download JSON ↗

codexgpt-5.6-sol684 turns57 narration627 reasoning627 action627 observation1 final

#476

narration

The repeat is past step 60 without divergence. I’m preserving the current final files as the validated method throughout this run; only evaluation artifacts and checkpoints are being added outside the submission.

#477

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:57840,chars:\"\",yield_time_ms:30000,max_output_tokens:80});\ntext(r.output);\n"
}

exec result

Script running with cell ID 250
Wall time 11.0 seconds
Output:

#478

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "250",
  "yield_time_ms": 30000,
  "max_tokens": 80
}

wait result

Script completed
Wall time 0.7 seconds
Output:
Warning: truncated output (original token count: 107)
Total output lines: 8

Warning: truncated output (original token count: 1069)
Total output lines: 37

vLLM generation done - elapsed time: 5.07s, prompts: 4, total tokens: 5162, avg l…27 tokens truncated…s_step_70.json
================================================================================


 71%|███████   | 71/100 [14:04<05:47, 11.99s/it]

#479

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:57840,chars:\"\",yield_time_ms:30000,max_output_tokens:80});\ntext(r.output);\n"
}

exec result

Script running with cell ID 251
Wall time 11.0 seconds
Output:

#480

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "251",
  "yield_time_ms": 30000,
  "max_tokens": 80
}

wait result

Script completed
Wall time 2.0 seconds
Output:
Warning: truncated output (original token count: 106)
Total output lines: 6

Warning: truncated output (original token count: 981)
Total output lines: 30

vLLM generation done - elapsed time: 4.99s, prompts: 4, total tokens: 4097, avg le…26 tokens truncated…90, avg length: 1522.5, speed: 1186.8 tok/s
vLLM generation done - elapsed time: 5.16s, prompts: 4, total tokens: 6144, avg length: 1536.0, speed: 1191.6 tok/s

#481

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:57840,chars:\"\",yield_time_ms:30000,max_output_tokens:80});\ntext(r.output);\n"
}

exec result

Script running with cell ID 252
Wall time 11.0 seconds
Output:

#482

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "252",
  "yield_time_ms": 30000,
  "max_tokens": 80
}

wait result

Script completed
Wall time 12.3 seconds
Output:
Warning: truncated output (original token count: 106)
Total output lines: 6

Warning: truncated output (original token count: 873)
Total output lines: 31

vLLM generation done - elapsed time: 5.09s, prompts: 4, total tokens: 5687, avg le…26 tokens truncated…93, avg length: 1498.2, speed: 1169.7 tok/s
vLLM generation done - elapsed time: 5.12s, prompts: 4, total tokens: 6051, avg length: 1512.8, speed: 1181.1 tok/s

#483

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:57840,chars:\"\",yield_time_ms:30000,max_output_tokens:80});\ntext(r.output);\n"
}

exec result

Script running with cell ID 253
Wall time 11.0 seconds
Output:

#484

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "253",
  "yield_time_ms": 30000,
  "max_tokens": 80
}

wait result

Script completed
Wall time 10.0 seconds
Output:
Warning: truncated output (original token count: 107)
Total output lines: 6

Warning: truncated output (original token count: 1186)
Total output lines: 40

vLLM generation done - elapsed time: 5.03s, prompts: 4, total tokens: 5006, avg l…27 tokens truncated…05, avg length: 1501.2, speed: 1170.8 tok/s
vLLM generation done - elapsed time: 5.14s, prompts: 4, total tokens: 6144, avg length: 1536.0, speed: 1196.0 tok/s

#485

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:57840,chars:\"\",yield_time_ms:30000,max_output_tokens:80});\ntext(r.output);\n"
}

exec result

Script running with cell ID 254
Wall time 11.0 seconds
Output:

#486

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "254",
  "yield_time_ms": 30000,
  "max_tokens": 80
}

wait result

Script completed
Wall time 12.8 seconds
Output:
Warning: truncated output (original token count: 106)
Total output lines: 6

Warning: truncated output (original token count: 870)
Total output lines: 26

vLLM generation done - elapsed time: 4.92s, prompts: 4, total tokens: 4358, avg le…26 tokens truncated…11, avg length: 1427.8, speed: 1116.5 tok/s
vLLM generation done - elapsed time: 5.13s, prompts: 4, total tokens: 6144, avg length: 1536.0, speed: 1198.2 tok/s

#487

narration

The repeat is above step 80 and still matches the first run’s profile. It should finish in a few minutes, followed by the full 30×12 evaluation.

#488

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:57840,chars:\"\",yield_time_ms:30000,max_output_tokens:80});\ntext(r.output);\n"
}

exec result

Script running with cell ID 255
Wall time 11.0 seconds
Output:

#489

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "255",
  "yield_time_ms": 30000,
  "max_tokens": 80
}

wait result

Script completed
Wall time 14.2 seconds
Output:
Warning: truncated output (original token count: 107)
Total output lines: 6

Warning: truncated output (original token count: 3709)
Total output lines: 235

vLLM generation done - elapsed time: 5.07s, prompts: 4, total tokens: 5135, avg …27 tokens truncated…44, avg length: 1536.0, speed: 1192.7 tok/s
vLLM generation done - elapsed time: 5.16s, prompts: 4, total tokens: 6143, avg length: 1535.8, speed: 1190.4 tok/s

#490

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:57840,chars:\"\",yield_time_ms:30000,max_output_tokens:80});\ntext(r.output);\n"
}

exec result

Script running with cell ID 256
Wall time 11.0 seconds
Output:

#491

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "256",
  "yield_time_ms": 30000,
  "max_tokens": 80
}

wait result

Script completed
Wall time 12.6 seconds
Output:
Warning: truncated output (original token count: 107)
Total output lines: 6

Warning: truncated output (original token count: 2394)
Total output lines: 279

vLLM generation done - elapsed time: 5.06s, prompts: 4, total tokens: 5504, avg …27 tokens truncated…42, avg length: 1410.5, speed: 1099.6 tok/s
vLLM generation done - elapsed time: 5.16s, prompts: 4, total tokens: 6144, avg length: 1536.0, speed: 1190.4 tok/s

#492

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:57840,chars:\"\",yield_time_ms:30000,max_output_tokens:80});\ntext(r.output);\n"
}

exec result

Script running with cell ID 257
Wall time 11.0 seconds
Output:

#493

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "257",
  "yield_time_ms": 30000,
  "max_tokens": 80
}

wait result

Script completed
Wall time 0.0 seconds
Output:
Warning: truncated output (original token count: 107)
Total output lines: 6

Warning: truncated output (original token count: 2221)
Total output lines: 241

vLLM generation done - elapsed time: 5.13s, prompts: 4, total tokens: 5911, avg …27 tokens truncated…09, avg length: 1502.2, speed: 1170.3 tok/s
vLLM generation done - elapsed time: 5.16s, prompts: 4, total tokens: 6144, avg length: 1536.0, speed: 1190.4 tok/s

#494

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:57840,chars:\"\",yield_time_ms:30000,max_output_tokens:80});\ntext(r.output);\n"
}

exec result

Script running with cell ID 258
Wall time 11.0 seconds
Output:

#495

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "258",
  "yield_time_ms": 30000,
  "max_tokens": 80
}

wait result

Script completed
Wall time 5.5 seconds
Output:
Warning: truncated output (original token count: 107)
Total output lines: 6

Warning: truncated output (original token count: 1077)
Total output lines: 37

vLLM generation done - elapsed time: 4.32s, prompts: 4, total tokens: 4103, avg l…27 tokens truncated…531174, 'learning_rate': 1.5000000000000002e-07, 'on_policy_loss': -0.0102, 'epoch': 0.86}

 98%|█████████▊| 98/100 [19:28<00:23, 11.98s/it]

#496

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:57840,chars:\"\",yield_time_ms:30000,max_output_tokens:800});\ntext(r.output);\n"
}

exec result

Script running with cell ID 259
Wall time 11.0 seconds
Output:

#497

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "259",
  "yield_time_ms": 30000,
  "max_tokens": 800
}

wait result

Script completed
Wall time 0.0 seconds
Output:
Warning: truncated output (original token count: 826)
Total output lines: 24

Warning: truncated output (original token count: 936)
Total output lines: 24

vLLM generation done - elapsed time: 4.29s, prompts: 4, total tokens: 4428, avg length: 1107.0, speed: 1031.7 tok/s
vLLM generation done - elapsed time: 5.07s, prompts: 4, total tokens: 5617, avg length: 1404.2, speed: 1107.6 tok/s
vLLM generation done - elapsed time: 5.13s, prompts: 4, total tokens: 5812, avg length: 1453.0, speed: 1134.0 tok/s
vLLM generation done - elapsed time: 5.13s, prompts: 4, total tokens: 6144, avg length: 1536.0, speed: 1196.7 tok/s
vLLM generation done - elapsed time: 4.90s, prompts: 4, total tokens: 4133, avg length: 1033.2, speed: 843.7 tok/s
vLLM generation done - elapsed time: 5.01s, prompts: 4, total tokens: 4595, avg length: 1148.8, speed: 916.8 tok/s
vLLM generation done - elapsed time: 5.12s, prompts: 4, total tokens: 5833, avg length: 1458.2, speed: 1138.7 tok/s
vLLM generation done - elapsed time: 5.15s, prompts: 4, total tokens: 6144, avg length: 1536.0, speed: 1194.1 tok/s

 99%|█████████▉| 99/100 [19:40<00:11, 11.97s/it]vLLM generation done - elapsed time: 5.08s, prompts: 4, total tokens: 5550, avg length: 1387.5, speed: 1092.1 tok/s
vLLM generation done - elapsed time: 5.11s, prompts: 4, total tokens: 5569, avg length: 1392.2, speed: 1090.8 tok/s
vLLM generation done - elapsed time: 5.12s, prompts: 4, total tokens: 5975, avg length: 1493.8, speed: 1166.6 tok/s
vLLM generation done - elapsed time: 5.13s, prompts: 4, total tokens: 5518, avg length: 1379.5, speed: 1075.2 tok/s
vLLM generation done - elapsed time: 5.07s, prompts: 4, total t…26 tokens truncated…olicy_loss': -0.0095, 'epoch': 0.87}

100%|██████████| 100/100 [19:52<00:00, 11.99s/it]
                                                 
{'train_runtime': 1195.3092, 'train_samples_per_second': 2.677, 'train_steps_per_second': 0.084, 'train_loss': -0.005127556845545769, 'epoch': 0.87}

100%|██████████| 100/100 [19:55<00:00, 11.99s/it]
100%|██████████| 100/100 [19:55<00:00, 11.95s/it]
[rank0]:[W802 20:47:11.704613587 ProcessGroupNCCL.cpp:1538] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
[rank1]:[W802 20:47:11.072961469 ProcessGroupNCCL.cpp:1538] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
[rank2]:[W802 20:47:11.124051222 ProcessGroupNCCL.cpp:1538] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
[rank3]:[W802 20:47:11.124195912 ProcessGroupNCCL.cpp:1538] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
TRAIN_CKPT=/workspace/runs/refined_repeat2/capped_seed42/checkpoint-100

#498

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.exec_command({cmd:\"bash /workspace/dev_eval.sh /workspace/runs/refined_repeat2/capped_seed42/checkpoint-100 /workspace/eval_results/refined_repeat2_full.json\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":300});\ntext(r.output); if(r.session_id) text(`SESSION_ID=${r.session_id}`);\n"
}

exec result

Script completed
Wall time 1.2 seconds
Output:
SESSION_ID=50469

#499

narration

The repeat finished all 100 steps with essentially the same aggregate training loss (−0.00513 versus −0.00505). Its full AIME evaluation is now running; this is a genuinely independent retrain despite the fixed nominal seed because vLLM rollouts and CUDA execution are nondeterministic.

#500

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:50469,chars:\"\",yield_time_ms:30000,max_output_tokens:150});\ntext(r.output);\n"
}

exec result

Script running with cell ID 261
Wall time 11.0 seconds
Output: