Safety Refusal on Jailbreak Prompts
87.2Refusal RateITI
Evaluation Results
| Method | Links | |
|---|---|---|
| ITIBackbone=Llama 3.2-1B2025.05 | 87.2 | |
| WASBackbone=Mistral 7B2025.05 | 81.7 | |
| WASBackbone=Llama 3.1-8B2025.05 | 78.9 | |
| ACTBackbone=Llama 3.2-1B2025.05 | 78.5 | |
| WASBackbone=Llama 3.2-1B2025.05 | 78 | |
| ITIBackbone=Llama 3.1-8B2025.05 | 68.9 | |
| ACTBackbone=Llama 3.1-8B2025.05 | 66 | |
| CASTBackbone=Mistral 7B2025.05 | 64.3 | |
| CASTBackbone=Llama 3.1-8B2025.05 | 63.9 | |
| CASTBackbone=Llama 3.2-1B2025.05 | 63.2 | |
| ACTBackbone=Mistral 7B2025.05 | 60.2 | |
| ITIBackbone=Mistral 7B2025.05 | 59 | |
| Base ModelBackbone=Mistral 7B2025.05 | 14 | |
| Base ModelBackbone=Llama 3.1-8B2025.05 | 12.2 | |
| Base ModelBackbone=Llama 3.2-1B2025.05 | 12 |