Targeted Refusal on Llama2-7B Generation Evaluation Set
93.14Completion Accuracy (CA)Vanilla
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VanillaAttack=CBA2025.11 | 93.14 | 89.36 | |
| VanillaAttack=VPI2025.11 | 91.25 | 97.36 | |
| VanillaAttack=Sleeper Agent2025.11 | 90.88 | 96.64 | |
| OURSAttack=VPI2025.11 | 90.1 | 14.45 | |
| GracCeFulAttack=CBA2025.11 | 89.37 | 20.95 | |
| CleanGenAttack=CBA2025.11 | 88.92 | 24.87 | |
| OURSAttack=CBA2025.11 | 88.49 | 15.4 | |
| OURSAttack=Sleeper Agent2025.11 | 88.14 | 11.37 | |
| CleanGenAttack=VPI2025.11 | 87.29 | 28.04 | |
| GracCeFulAttack=VPI2025.11 | 87.15 | 20.17 | |
| GracCeFulAttack=Sleeper Agent2025.11 | 86.78 | 21.45 | |
| CleanGenAttack=Sleeper Agent2025.11 | 86.34 | 25.8 | |
| MuScleLoRAAttack=VPI2025.11 | 83.15 | 19.57 | |
| MuScleLoRAAttack=CBA2025.11 | 82.37 | 19.2 | |
| MuScleLoRAAttack=Sleeper Agent2025.11 | 80.45 | 16.44 |