Output-based feature description evaluation on Gemma-2 MLP SAE features
49.9ScoreEnsemble Concat
Evaluation Results
| Method | Links | |
|---|---|---|
| Ensemble Concatcombination=All2025.01 | 49.9 | |
| Ensemble Rawcombination=All2025.01 | 48.7 | |
| Ensemble Rawcombination=VP+TC2025.01 | 46.6 | |
| Ensemble Rawcombination=MA+TC2025.01 | 46.2 | |
| TokenChange2025.01 | 40.3 | |
| Ensemble Rawcombination=MA+VP2025.01 | 38.6 | |
| VocabProj2025.01 | 37.2 | |
| MaxAct2025.01 | 35.1 |