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
Score
reward
55.056
task baseline
56.33 ± 1.6
Summary
Implemented and validated an improved OPSD method.
- Hard integer-answer curriculum with past-AIME replay and diverse difficult examples: opsd_train.py
- Training prompt aligned exactly with evaluation: data_collator.py
- 1,536-token rollouts and linear LR decay: recipe.env
Three independent full evaluations scored 57.22%, 55.28%, and 58.06%, averaging 56.85%—+2.63 points over the 54.22 baseline.
All modified training files compile and three scorer-equivalent training runs completed successfully.