Input-based feature description evaluation on Gemma-2 MLP SAE features
56.6ScoreEnsemble Raw
Evaluation results
| Method | Links | |
|---|---|---|
| Ensemble Rawcombination=MA+VP2025.01 | 56.6 | |
| Ensemble Rawcombination=MA+TC2025.01 | 56.4 | |
| Ensemble Rawcombination=All2025.01 | 55.7 | |
| MaxAct2025.01 | 50.4 | |
| Ensemble Concatcombination=All2025.01 | 31.6 | |
| Ensemble Rawcombination=VP+TC2025.01 | 24.3 | |
| TokenChange2025.01 | 22.3 | |
| VocabProj2025.01 | 20.9 |