Sparse Autoencoding on Mistral-7B activations
95.3KL ScoreSASA
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SASATraining Token Budget=500M2026.06 | 95.3 | 95.1 | 98.6 | 80 | |
| Standard SAETraining Token Budget=1B2026.06 | 93.2 | 94.1 | 96.7 | 285 |
| Method | Links | ||||
|---|---|---|---|---|---|
| SASATraining Token Budget=500M2026.06 | 95.3 | 95.1 | 98.6 | 80 | |
| Standard SAETraining Token Budget=1B2026.06 | 93.2 | 94.1 | 96.7 | 285 |