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
- 428
- duration
- 9h 21m
- recorded
- 2026-08-02
run idopsd-aime24-4gpus__eNDoWrD
Score
reward
56.556
task baseline
56.33 ± 1.6
Summary
Implemented a reasoning-effort curriculum in opsd_train.py:
- Trains only on dataset examples with
generated_token_count >= 3000—13,565 hard, diverse problems. - Uses 2,048-token rollouts to cover later reasoning and answers.
- Retains the paper’s linear schedule and original OPSD loss.
Full AIME24 dev results:
- 1,024-token retrains: 55.56%, 57.22%
- Final 2,048-token method: 57.22%
- Baseline anchor: 54.22%
Final settings are in recipe.env. Syntax and scorer compatibility were verified.