Safety Evaluation on HarmBench Direct
99.75Safety Score (1-ASR)RESTA
Evaluation Results
| Method | Links | |
|---|---|---|
| RESTABackbone Model=Lingshu-7B, Alignment Method=RESTA2025.12 | 99.75 | |
| PSIBackbone Model=Lingshu-7B, Alignment Method=Parameter-Space Intervention2025.12 | 90.5 | |
| PSIBackbone Model=HuatuoGPT-Vision-7B, Alignment Method=Parameter-Space Intervention2025.12 | 89.5 | |
| RESTABackbone Model=HuatuoGPT-Vision-7B, Alignment Method=RESTA2025.12 | 85.75 | |
| ModelMergeBackbone Model=Lingshu-7B, Alignment Method=ModelMerge2025.12 | 61 | |
| OriginalBackbone Model=Lingshu-7B, Alignment Method=None2025.12 | 47.5 | |
| ModelMergeBackbone Model=HuatuoGPT-Vision-7B, Alignment Method=ModelMerge2025.12 | 38.75 | |
| OriginalBackbone Model=HuatuoGPT-Vision-7B, Alignment Method=None2025.12 | 29.5 |