Sparse Autoencoding on gemma2-2b-layer-13 activations
100.7L0HSAE
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| HSAESparsity level=L0 ~ 1002026.02 | 100.7 | 0.757 | 0.747 | 0.415 | 46.3 | |
| BaselineSparsity level=L0 ~ 1002026.02 | 100 | 0.758 | 0.619 | 0.317 | 78.9 | |
| HSAESparsity level=L0 ~ 802026.02 | 80.1 | 0.743 | 0.749 | 0.421 | 36.2 | |
| BaselineSparsity level=L0 ~ 802026.02 | 80 | 0.743 | 0.634 | 0.326 | 61.9 | |
| HSAESparsity level=L0 ~ 502026.02 | 50.1 | 0.713 | 0.769 | 0.432 | 21.3 | |
| BaselineSparsity level=L0 ~ 502026.02 | 49.8 | 0.714 | 0.662 | 0.343 | 36.1 |