Safety Evaluation on Q-LatHarmful
6.52Attack Success Rate (ASR)Qwen-2.5-7B-Instruct
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Qwen-2.5-7B-InstructBackbone=Qwen-2.5-7B-Instruct2026.05 | 6.52 | 1.22 | |
| LisaBackbone=Qwen-2.5-7B-Instruct2026.05 | 7.33 | 1.24 | |
| SPARDBackbone=Qwen-2.5-7B-Instruct2026.05 | 7.94 | 1.27 | |
| LisaBase Model=LLaMA-3.2-3B-Instruct, Defense Method=Lisa2026.05 | 8.96 | 1.29 | |
| SPARDBase Model=LLaMA-3.2-3B-Instruct, Defense Method=SPARD2026.05 | 9.16 | 1.35 | |
| SafeGradBackbone=Qwen-2.5-7B-Instruct2026.05 | 9.78 | 1.32 | |
| LLaMA-3.2-3B-InstructBase Model=LLaMA-3.2-3B-Instruct2026.05 | 12.02 | 1.4 | |
| PTSTBackbone=Qwen-2.5-7B-Instruct2026.05 | 16.9 | 1.59 | |
| SFTBackbone=Qwen-2.5-7B-Instruct2026.05 | 28.92 | 2.05 | |
| SafeInstrBackbone=Qwen-2.5-7B-Instruct2026.05 | 29.53 | 2.11 | |
| SafeInstrBase Model=LLaMA-3.2-3B-Instruct, Defense Method=SafeInstr2026.05 | 89.21 | 4.52 | |
| PTSTBase Model=LLaMA-3.2-3B-Instruct, Defense Method=PTST2026.05 | 96.33 | 4.73 | |
| SafeGradBase Model=LLaMA-3.2-3B-Instruct, Defense Method=SafeGrad2026.05 | 98.37 | 4.86 | |
| SFTBase Model=LLaMA-3.2-3B-Instruct, Defense Method=SFT2026.05 | 99.8 | 4.95 |