Refusal Intervention Analysis on Refusal Dataset
-2.14Refusal Score (Policy)Llama-3-8B-Instruct
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Llama-3-8B-InstructLayer=-8, Alpha (α)=1.02026.06 | -2.14 | -1.88 | -4.25 | -4.87 | -3.66 | |
| Gemma-2-9B-ITLayer=-8, Alpha (α)=1.02026.06 | -8.17 | 3.84 | -7.86 | -0.47 | 0.98 | |
| Mistral-7B-Instruct-v0.3Layer=-8, Alpha (α)=1.02026.06 | -16.2 | -1.48 | -0.57 | -0.15 | -17.27 |