Concept Steering on GPT-2 (3 concepts x 4 layers x 25 contexts) (test)
1.32Median Off-target KL RatioEuclidean
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| EuclideanTarget concept probability (P^W)=0.92026.05 | 1.32 | 72.9 | 10 | 96 | |
| CAATarget concept probability (P^W)=0.52026.05 | 1.42 | 77.6 | 10 | 76 | |
| Strongest per-caseTarget concept probability (P^W)=0.52026.05 | 1.44 | 78.3 | 10 | 83 | |
| CAATarget concept probability (P^W)=0.92026.05 | 1.5 | 75.8 | 10 | 66 | |
| Strongest per-caseTarget concept probability (P^W)=0.92026.05 | 1.5 | 72.2 | 1 | 72 | |
| Strongest per-caseTarget concept probability (P^W)=0.72026.05 | 1.52 | 74 | 10 | 77 | |
| CAATarget concept probability (P^W)=0.72026.05 | 1.54 | 76.1 | 10 | 71 | |
| CAATarget concept probability (P^W)=0.32026.05 | 1.58 | 78 | 10 | 82 | |
| EuclideanTarget concept probability (P^W)=0.72026.05 | 1.61 | 78.4 | 10 | 102 | |
| Strongest per-caseTarget concept probability (P^W)=0.32026.05 | 1.65 | 78.9 | 10 | 90 | |
| EuclideanTarget concept probability (P^W)=0.52026.05 | 1.91 | 80.6 | 10 | 108 | |
| EuclideanTarget concept probability (P^W)=0.32026.05 | 2.51 | 81.4 | 10 | 113 |