Verse Style-Transfer (held-in and held-out)
46Activation PatchingDifference-in-means
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Difference-in-meansModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.02 | 46 | 46 | 21 | 15 | 13 | |
| Difference-in-meansModel=OLMo-13B, alpha=10, k=12, N=1202026.02 | 34 | 36 | 20 | 12 | 11 | |
| ReFTModel=OLMo-13B, alpha=10, k=12, N=1202026.02 | 33 | 45 | 39 | 25 | 43 | |
| Difference-in-meansModel=Qwen-14B, alpha=10, k=12, N=1202026.02 | 28 | 30 | 36 | 17 | 17 | |
| ReFTModel=Qwen-14B, alpha=10, k=12, N=1202026.02 | 24 | 27 | 7 | 26 | 13 | |
| Mean steeringModel=Qwen-14B, alpha=10, k=12, N=1202026.02 | 20 | 18 | 8 | 45 | 10 | |
| Mean steeringModel=OLMo-13B, alpha=10, k=12, N=1202026.02 | 16 | 18 | 9 | 11 | 14 | |
| Mean steeringModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.02 | 13 | 10 | 2 | 33 | 5 | |
| ReFTModel=SOLAR-10.7B, alpha=10, k=12, N=1202026.02 | 0 | 0 | 0 | 0 | 0 |