Sparse Autoencoder Evaluation on SAEBench Pythia-70M
0.308RAVEL ScoreEnsembling (Boosting)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Ensembling (Boosting)Backbone=Pythia-70M, Ensemble Size=8 SAEs, Number of Runs=5, Strategy=Boosting2025.05 | 0.308 | 0.9974 | 0.0187 | |
| Base SAEBackbone=Pythia-70M, Number of Runs=52025.05 | 0.3078 | 0.9822 | 0.0164 | |
| Ensembling (NB)Backbone=Pythia-70M, Ensemble Size=8 SAEs, Number of Runs=5, Strategy=Naive Bagging2025.05 | 0.3078 | 0.9949 | 0.0075 | |
| Expanded SAEBackbone=Pythia-70M, Number of Runs=52025.05 | 0.3077 | 0.9845 | 0.0147 |