Jailbreak Attack on AdvBench Behaviours
53.8ASRGCG
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| GCGTarget Model=Llama2-chat-7B2024.05 | 53.8 | 1 | 20.6 | 53.3 | |
| ADCTarget Model=Llama2-chat-7B2024.05 | 96.2 | 0.93 | 11.1 | 95.8 | |
| ADC+Target Model=Llama2-chat-7B2024.05 | 96.5 | 0.67 | 8.2 | 96.2 | |
| GCGTarget Model=Zephyr-β-7B2024.05 | 98.1 | 1 | 13.2 | 68.3 | |
| ADC+Target Model=Zephyr-β-7B2024.05 | 98.8 | 0.67 | 6.1 | 97.5 | |
| GCGTarget Model=Vicuna-v1.5-7B2024.05 | 99.4 | 1 | 3 | 98.3 | |
| ADCTarget Model=Vicuna-v1.5-7B2024.05 | 99.8 | 0.93 | 2.3 | 99.8 | |
| ADC+Target Model=Vicuna-v1.5-7B2024.05 | 99.8 | 0.67 | 1.6 | 99.8 | |
| ADCTarget Model=Zephyr-β-7B2024.05 | 99.8 | 0.93 | 7.9 | 92.1 |