Refusal Induction on Refusal Induction held-in and held-out
55Activation PatchingDifference-in-means
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Difference-in-meansModel=OLMo-13B, alpha=10, k=12, N=1202026.02 | 55 | 55 | 41 | 58 | 28 | |
| Mean steeringModel=Qwen-14B, alpha=10, k=12, N=1202026.02 | 48 | 45 | 8 | 18 | 20 | |
| Mean steeringModel=OLMo-13B, alpha=10, k=12, N=1202026.02 | 48 | 49 | 22 | 34 | 30 | |
| Difference-in-meansModel=Qwen-14B, alpha=10, k=12, N=1202026.02 | 46 | 41 | 27 | 20 | 25 | |
| ReFTModel=Qwen-14B, alpha=10, k=12, N=1202026.02 | 40 | 34 | 6 | 23 | 5 | |
| ReFTModel=OLMo-13B, alpha=10, k=12, N=1202026.02 | 24 | 24 | 14 | 31 | 27 | |
| Difference-in-meansModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.02 | 24 | 23 | 20 | 14 | 13 | |
| Mean steeringModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.02 | 12 | 10 | 12 | 6 | 6 | |
| ReFTModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.02 | 10 | 8 | 10 | 5 | 5 |