Output-based feature description evaluation on Llama Residual SAE features 3.1
75.4ScoreEnsemble Concat
Evaluation Results
| Method | Links | |
|---|---|---|
| Ensemble Concatcombination=All2025.01 | 75.4 | |
| TokenChange2025.01 | 72 | |
| MaxAct2025.01 | 71.8 | |
| Ensemble Rawcombination=All2025.01 | 71.2 | |
| Ensemble Rawcombination=MA+VP2025.01 | 68.9 | |
| Ensemble Rawcombination=MA+TC2025.01 | 68 | |
| Ensemble Rawcombination=VP+TC2025.01 | 67.4 | |
| VocabProj2025.01 | 64.2 |