Prompt-side safety risk classification on Sorry
84.6F1 ScoreStage3-DPO
Evaluation Results
| Method | Links | |
|---|---|---|
| Stage3-DPOParams=4B2026.07 | 84.6 | |
| Stage2-SFTParams=4B2026.07 | 81.8 | |
| Stage1-SFT-v0Params=4B2026.07 | 81.6 | |
| Stage1-SFT-v1Params=4B2026.07 | 79.8 | |
| Qwen3Guard-8B-GenParams=8B2026.07 | 79.7 | |
| YuFeng-XGuard-Reason-8BParams=8B2026.07 | 77.6 | |
| Stage1-SFT-v3Params=4B2026.07 | 77.2 | |
| Stage1-SFT-v2Params=4B2026.07 | 75.7 | |
| Qwen3Guard-4B-GenParams=4B2026.07 | 74.2 | |
| Qwen3Guard-0.8B-GenParams=0.8B2026.07 | 71.1 | |
| Stage1-SFT-v4Params=4B2026.07 | 70.2 | |
| YuFeng-XGuard-Reason-0.6BParams=0.6B2026.07 | 65.8 |