Safety Evaluation on I-BeaverTails
14.68Attack Success Rate (ASR)SPARD
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SPARDBase Model=LLaMA-3.2-3B-Instruct, Defense Method=SPARD2026.05 | 14.68 | 1.42 | |
| SPARDBackbone=Qwen-2.5-7B-Instruct2026.05 | 20.13 | 1.57 | |
| LisaBackbone=Qwen-2.5-7B-Instruct2026.05 | 34.8 | 1.93 | |
| Qwen-2.5-7B-InstructBackbone=Qwen-2.5-7B-Instruct2026.05 | 36.06 | 1.97 | |
| SafeGradBackbone=Qwen-2.5-7B-Instruct2026.05 | 36.29 | 2.01 | |
| LisaBase Model=LLaMA-3.2-3B-Instruct, Defense Method=Lisa2026.05 | 45.7 | 2.32 | |
| LLaMA-3.2-3B-InstructBase Model=LLaMA-3.2-3B-Instruct2026.05 | 52.2 | 2.44 | |
| SFTBackbone=Qwen-2.5-7B-Instruct2026.05 | 56.39 | 2.81 | |
| PTSTBackbone=Qwen-2.5-7B-Instruct2026.05 | 56.39 | 2.81 | |
| SafeGradBase Model=LLaMA-3.2-3B-Instruct, Defense Method=SafeGrad2026.05 | 59.12 | 2.77 | |
| SafeInstrBackbone=Qwen-2.5-7B-Instruct2026.05 | 59.75 | 2.95 | |
| PTSTBase Model=LLaMA-3.2-3B-Instruct, Defense Method=PTST2026.05 | 64.99 | 3.04 | |
| SafeInstrBase Model=LLaMA-3.2-3B-Instruct, Defense Method=SafeInstr2026.05 | 72.54 | 3.46 | |
| SFTBase Model=LLaMA-3.2-3B-Instruct, Defense Method=SFT2026.05 | 79.45 | 3.66 |