Instruction Following Evaluation on HelpSteer2 (val)
71.8QualitySEE
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SEEJudge=Gemini 3.1 Flash-Lite2026.06 | 71.8 | 70.9 | — | — | |
| SEEunique examples=160, batch size=16, rounds=152026.06 | 70.4 | 73.1 | 67.1 | 70 | |
| Adapted RLCRJudge=Gemini 3.1 Flash-Lite2026.06 | 69.1 | 67 | — | — | |
| Qwen 3-4BJudge=Gemini 3.1 Flash-Lite2026.06 | 68.2 | 65.1 | — | — | |
| Adapted RLCRunique examples=~5,000, batch size=482026.06 | 66.2 | 67.5 | 57 | 61.7 | |
| Qwen3-4Bstatus=base model2026.06 | 64.4 | 63.2 | 50 | 50 | |
| SEEJudge=Claude Sonnet 4.62026.06 | 59.9 | 67 | — | — | |
| Adapted RLCRJudge=Claude Sonnet 4.62026.06 | 58.1 | 62.1 | — | — | |
| Qwen 3-4BJudge=Claude Sonnet 4.62026.06 | 57.5 | 58.5 | — | — |