Safety Refusal on ToxicChat
95Refusal RateWAS
Evaluation Results
| Method | Links | |
|---|---|---|
| WASBackbone=Mistral 7B2025.05 | 95 | |
| WASBackbone=Llama 3.1-8B2025.05 | 93 | |
| WASBackbone=Llama 3.2-1B2025.05 | 91 | |
| ITIBackbone=Llama 3.2-1B2025.05 | 89.4 | |
| ACTBackbone=Llama 3.2-1B2025.05 | 59.8 | |
| ITIBackbone=Llama 3.1-8B2025.05 | 49.6 | |
| CASTBackbone=Llama 3.1-8B2025.05 | 46 | |
| CASTBackbone=Mistral 7B2025.05 | 43.3 | |
| CASTBackbone=Llama 3.2-1B2025.05 | 39.2 | |
| Base ModelBackbone=Llama 3.1-8B2025.05 | 32 | |
| ACTBackbone=Llama 3.1-8B2025.05 | 30 | |
| ACTBackbone=Mistral 7B2025.05 | 29.5 | |
| Base ModelBackbone=Llama 3.2-1B2025.05 | 29 | |
| ITIBackbone=Mistral 7B2025.05 | 27.3 | |
| Base ModelBackbone=Mistral 7B2025.05 | 27 |