Input-based feature description evaluation on Llama Residual SAE features 3.1
37.2ScoreEnsemble Raw
Evaluation Results
| Method | Links | |
|---|---|---|
| Ensemble Rawcombination=MA+TC2025.01 | 37.2 | |
| Ensemble Rawcombination=MA+VP2025.01 | 36.9 | |
| Ensemble Rawcombination=All2025.01 | 36 | |
| MaxAct2025.01 | 30.3 | |
| Ensemble Concatcombination=All2025.01 | 28.5 | |
| TokenChange2025.01 | 21.4 | |
| Ensemble Rawcombination=VP+TC2025.01 | 20.9 | |
| VocabProj2025.01 | 18.2 |