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

#376

narration

The robust-KL/2,048-token checkpoint completed all 100 steps with a positive final loss (0.0116) and is now in the full official evaluation. The validated 57.22% configuration remains the fallback, so this experiment can only improve the final selection, not displace it on training aesthetics alone.

#377

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 203
Wall time 11.0 seconds
Output:

#378

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "203",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

wait result

Script completed
Wall time 11.6 seconds
Output:

#379

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 204
Wall time 11.0 seconds
Output:

#380

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "204",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

wait result

Script completed
Wall time 14.0 seconds
Output:

#381

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 205
Wall time 11.0 seconds
Output:

#382

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "205",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

wait result

Script completed
Wall time 12.0 seconds
Output:

#383

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 206
Wall time 11.0 seconds
Output:

#384

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "206",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

wait result

Script completed
Wall time 13.7 seconds
Output:

Processed prompts:  12%|█▎        | 12/96 [02:48<19:38, 14.03s/it, est. speed input: 11.12 toks/s, output: 452.29 toks/s]

#385

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 207
Wall time 11.0 seconds
Output:

#386

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "207",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

wait result

Script completed
Wall time 11.6 seconds
Output:

#387

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 208
Wall time 11.0 seconds
Output:

#388

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "208",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

wait result

Script completed
Wall time 11.3 seconds
Output:

Processed prompts:  25%|██▌       | 24/96 [03:25<09:05,  7.57s/it, est. speed input: 18.91 toks/s, output: 837.39 toks/s]
Processed prompts:  12%|█▎        | 12/96 [03:26<24:03, 17.19s/it, est. speed input: 8.61 toks/s, output: 446.56 toks/s]

#389

narration

Evaluation is progressing normally across all four shards. I’m holding the repository in the robust-KL state until the 360-sample result is known; reverting to the validated clipped-loss variant is a small, explicit patch if needed.

#390

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 209
Wall time 11.0 seconds
Output:

#391

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "209",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

wait result

Script completed
Wall time 14.0 seconds
Output:

#392

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 210
Wall time 11.0 seconds
Output:

#393

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "210",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

wait result

Script completed
Wall time 13.8 seconds
Output:

Processed prompts:  14%|█▍        | 12/84 [04:48<28:53, 24.08s/it, est. speed input: 3.45 toks/s, output: 425.26 toks/s]

#394

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 211
Wall time 11.0 seconds
Output:

#395

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "211",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

wait result

Script completed
Wall time 11.0 seconds
Output:

Processed prompts:  29%|██▊       | 24/84 [05:22<11:33, 11.56s/it, est. speed input: 5.95 toks/s, output: 743.28 toks/s]
Processed prompts:  43%|████▎     | 36/84 [05:47<05:46,  7.23s/it, est. speed input: 8.05 toks/s, output: 1037.82 toks/s]

#396

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 212
Wall time 11.0 seconds
Output:

#397

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "212",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

wait result

Script completed
Wall time 13.6 seconds
Output:

Processed prompts:  57%|█████▋    | 48/84 [06:10<03:04,  5.12s/it, est. speed input: 12.81 toks/s, output: 1420.50 toks/s]
Processed prompts:  38%|███▊      | 36/96 [06:32<11:14, 11.25s/it, est. speed input: 12.57 toks/s, output: 857.38 toks/s]

#398

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 213
Wall time 11.0 seconds
Output:

#399

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolwait

args

{
  "cell_id": "213",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

wait result

Script completed
Wall time 8.1 seconds
Output:

Processed prompts:  14%|█▍        | 12/84 [06:59<41:55, 34.94s/it, est. speed input: 2.40 toks/s, output: 395.27 toks/s]

#400

reasoning (redacted)

encrypted chain-of-thought — recorded, not exposed

toolexec

args

{
  "input": "const r = await tools.write_stdin({session_id:79122,chars:\"\",yield_time_ms:30000,max_output_tokens:300});\ntext(r.output);\n"
}

exec result

Script running with cell ID 214
Wall time 11.0 seconds
Output: