Instruction Following on IFBench (IFBench, Avg.)
23.88IFBench ScoreSFT Warm-up (With think)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SFT Warm-up (With think)Backbone=Qwen3-4B, Category=Rubric-as-Reward, Thinking Mode=With think2026.05 | 23.88 | 36.34 | |
| SFT Warm-upBackbone=Qwen3-4B, Category=Think-with-Rubrics (Ours)2026.05 | 22.79 | 39.12 | |
| + DAPO (Self-gen rubric only)Backbone=Qwen3-4B, Category=Think-with-Rubrics (Ours), Training Stage=DAPO2026.05 | 22.45 | 43.21 | |
| + DAPO (Mixed rubric reward)Backbone=Qwen3-4B, Category=Think-with-Rubrics (Ours), Training Stage=DAPO2026.05 | 21.77 | 41.48 | |
| Qwen3-4B-BaseBackbone=Qwen3-4B, Category=Base Model2026.05 | 21.43 | 29.2 | |
| SFT Warm-up (Direct answer)Backbone=Qwen3-4B, Category=Rubric-as-Reward, Thinking Mode=Direct answer2026.05 | 21.43 | 35.12 | |
| + DAPO (Golden rubric reward)Backbone=Qwen3-4B, Category=Rubric-as-Reward, Thinking Mode=With think, Training Stage=DAPO2026.05 | 20.07 | 39.05 | |
| + DAPO (Golden rubric reward)Backbone=Qwen3-4B, Category=Rubric-as-Reward, Thinking Mode=Direct answer, Training Stage=DAPO2026.05 | 19.05 | 37.72 | |
| + DAPO (Golden rubric only)Backbone=Qwen3-4B, Category=Think-with-Rubrics (Ours), Training Stage=DAPO2026.05 | 17.35 | 39.55 |