Held-out Skill Evaluation on Multi-family n=225 (holdout)
43.5556Solve RateGRPO from SFT init
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPO from SFT initbatches=40, initialization=SFT init2026.05 | 43.5556 | |
| GRPO from base initbatches=40, initialization=base init2026.05 | 27.5556 | |
| Base Qwen3-8B2026.05 | 11.5556 |