Safety Evaluation on Refusal Signal Score
7.5ASRMLP
Evaluation Results
| Method | Links | |
|---|---|---|
| MLPBase Model=Llama-3.1-70B, Condition=Cured2026.05 | 7.5 | |
| MLPBase Model=Qwen2.5-72B, Condition=Cured2026.05 | 8.06 | |
| MoEBase Model=Llama-3.1-8B, Condition=Cured2026.05 | 8.62 | |
| MoEBase Model=Gemma-2-27B, Condition=Cured2026.05 | 10.75 | |
| MoEBase Model=Gemma-2-9b, Condition=Cured2026.05 | 11.31 | |
| MoEBase Model=Mistral-7B-v0.3, Condition=Cured2026.05 | 13.31 | |
| MLPBase Model=Mistral-7B-v0.3, Condition=Cured2026.05 | 13.88 | |
| MLPBase Model=Qwen2.5-7B, Condition=Cured2026.05 | 14.12 | |
| MLPBase Model=Gemma-2-27B, Condition=Cured2026.05 | 16.06 | |
| MoEBase Model=Qwen2.5-7B, Condition=Cured2026.05 | 16.38 | |
| FMBase Model=Gemma-2-9b, Condition=Cured2026.05 | 20.12 | |
| FMBase Model=Gemma-2-27B, Condition=Cured2026.05 | 23.31 | |
| MLPBase Model=Llama-3.1-8B, Condition=Cured2026.05 | 24.06 | |
| MoEBase Model=Llama-3.1-70B, Condition=Cured2026.05 | 24.5 | |
| MLPBase Model=Gemma-2-9b, Condition=Cured2026.05 | 28.75 | |
| FMBase Model=Llama-3.1-8B, Condition=Cured2026.05 | 29.12 | |
| MoEBase Model=Qwen2.5-72B, Condition=Cured2026.05 | 29.25 | |
| CVAEBase Model=Qwen2.5-72B, Condition=Cured2026.05 | 29.75 | |
| FMBase Model=Qwen2.5-7B, Condition=Cured2026.05 | 37.31 | |
| FMBase Model=Mistral-7B-v0.3, Condition=Cured2026.05 | 45 | |
| Safe LoRABase Model=Qwen2.5-72B, Condition=Cured2026.05 | 46.38 | |
| CVAEBase Model=Gemma-2-9b, Condition=Cured2026.05 | 46.88 | |
| CVAEBase Model=Gemma-2-27B, Condition=Cured2026.05 | 48.19 | |
| AEBase Model=Gemma-2-27B, Condition=Cured2026.05 | 49.38 | |
| FMBase Model=Llama-3.1-70B, Condition=Cured2026.05 | 55.38 | |
| Safe LoRABase Model=Gemma-2-27B, Condition=Cured2026.05 | 56.75 | |
| CVAEBase Model=Qwen2.5-7B, Condition=Cured2026.05 | 58.19 | |
| CVAEBase Model=Llama-3.1-70B, Condition=Cured2026.05 | 58.56 | |
| CVAEBase Model=Mistral-7B-v0.3, Condition=Cured2026.05 | 58.75 | |
| AEBase Model=Gemma-2-9b, Condition=Cured2026.05 | 59.5 | |
| FMBase Model=Qwen2.5-72B, Condition=Cured2026.05 | 59.5 | |
| Safe LoRABase Model=Gemma-2-9b, Condition=Cured2026.05 | 59.62 | |
| AEBase Model=Qwen2.5-7B, Condition=Cured2026.05 | 60.75 | |
| Safe LoRABase Model=Mistral-7B-v0.3, Condition=Cured2026.05 | 64.44 | |
| Safe LoRABase Model=Qwen2.5-7B, Condition=Cured2026.05 | 67.19 | |
| AEBase Model=Qwen2.5-72B, Condition=Cured2026.05 | 67.75 | |
| Safe LoRABase Model=Llama-3.1-70B, Condition=Cured2026.05 | 67.94 | |
| CVAEBase Model=Llama-3.1-8B, Condition=Cured2026.05 | 68.19 | |
| AEBase Model=Llama-3.1-8B, Condition=Cured2026.05 | 72.06 | |
| AEBase Model=Mistral-7B-v0.3, Condition=Cured2026.05 | 72.31 | |
| Safe LoRABase Model=Llama-3.1-8B, Condition=Cured2026.05 | 72.62 | |
| AEBase Model=Llama-3.1-70B, Condition=Cured2026.05 | 75.38 | |
| Unsafe BaselineBase Model=Qwen2.5-72B, Condition=Unsafe (Finetuned)2026.05 | 80.5 | |
| SaLoRABase Model=Qwen2.5-72B, Condition=Cured2026.05 | 81.38 | |
| Unsafe BaselineBase Model=Qwen2.5-7B, Condition=Unsafe (Finetuned)2026.05 | 81.69 | |
| SaLoRABase Model=Qwen2.5-7B, Condition=Cured2026.05 | 81.94 | |
| Unsafe BaselineBase Model=Llama-3.1-70B, Condition=Unsafe (Finetuned)2026.05 | 86.38 | |
| SaLoRABase Model=Llama-3.1-70B, Condition=Cured2026.05 | 86.62 | |
| SaLoRABase Model=Gemma-2-27B, Condition=Cured2026.05 | 86.62 | |
| Unsafe BaselineBase Model=Gemma-2-27B, Condition=Unsafe (Finetuned)2026.05 | 86.94 | |
| Unsafe BaselineBase Model=Gemma-2-9b, Condition=Unsafe (Finetuned)2026.05 | 89.88 | |
| SaLoRABase Model=Gemma-2-9b, Condition=Cured2026.05 | 90.12 | |
| Unsafe BaselineBase Model=Llama-3.1-8B, Condition=Unsafe (Finetuned)2026.05 | 90.44 | |
| SaLoRABase Model=Llama-3.1-8B, Condition=Cured2026.05 | 90.88 | |
| SaLoRABase Model=Mistral-7B-v0.3, Condition=Cured2026.05 | 92 | |
| Unsafe BaselineBase Model=Mistral-7B-v0.3, Condition=Unsafe (Finetuned)2026.05 | 92.06 |