Harmful Request Compliance on Harmful Requests RepE Attack
91.2Compliance RateLlama-3-8B-Instruct
Evaluation Results
| Method | Links | |
|---|---|---|
| Llama-3-8B-InstructBase Model=Llama-3-8B-Instruct, Training Configuration=Refusal Trained2024.06 | 91.2 | |
| Mistral-7B-Instruct-v2Base Model=Mistral-7B-Instruct-v2, Training Configuration=Refusal Trained2024.06 | 73.7 | |
| R2D2Base Model=Mistral-7B-Instruct-v2, Training Configuration=Adv Trained2024.06 | 30.7 | |
| Representation ReroutingBase Model=Llama-3-8B-Instruct, Training Configuration=+ RR2024.06 | 8.7 | |
| Representation ReroutingBase Model=Mistral-7B-Instruct-v2, Training Configuration=+ RR2024.06 | 6.2 | |
| CygnetBase Model=Llama-3-8B-Instruct, Training Configuration=Circuit Breaker + RepE Fine-tune2024.06 | 0 |