Feature Interpretability on LLM Activations
86.9AutoInterp ScoreHSAE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| HSAEDictionary Size=163842026.02 | 86.9 | 32.4 | |
| BaselineDictionary Size=163842026.02 | 86.1 | 33.8 | |
| HSAEDictionary Size=81922026.02 | 85.9 | 31.4 | |
| BaselineDictionary Size=81922026.02 | 85.4 | 32 | |
| BaselineDictionary Size=40962026.02 | 82.9 | 30.2 | |
| HSAEDictionary Size=40962026.02 | 81.4 | 28.6 | |
| BaselineDictionary Size=20482026.02 | 81.3 | 24.5 | |
| HSAEDictionary Size=20482026.02 | 80.7 | 27.2 |