Sparse Autoencoder Evaluation on SAEBench Pythia-160M (test)
0.502RAVEL ScoreEnsembling (Boosting)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Ensembling (Boosting)Language Model=Pythia-160M, SAE Architecture=TopK, Activation Layer Used=8, Number of SAEs in Ensemble=8, Evaluation Runs=52025.05 | 0.502 | 1 | 0.2701 | |
| Ensembling (NB)Language Model=Pythia-160M, SAE Architecture=TopK, Activation Layer Used=8, Number of SAEs in Ensemble=8, Evaluation Runs=52025.05 | 0.4977 | 0.9829 | 0.0163 | |
| Expanded SAELanguage Model=Pythia-160M, SAE Architecture=TopK, Activation Layer Used=8, Activation Dimension=768, Evaluation Runs=52025.05 | 0.497 | 0.9783 | 0.2886 | |
| Base SAELanguage Model=Pythia-160M, SAE Architecture=TopK, Activation Layer Used=8, Activation Dimension=768, Evaluation Runs=52025.05 | 0.4969 | 0.9771 | 0.2635 |