Update README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# DINO Sparse Autoencoder Checkpoint
|
| 2 |
+
|
| 3 |
+
This model is a sparse autoencoder trained on DINO's internal representations.
|
| 4 |
+
|
| 5 |
+
## Model Details
|
| 6 |
+
|
| 7 |
+
### Architecture
|
| 8 |
+
- **Layer**: 2
|
| 9 |
+
- **Layer Type**: hook_resid_post
|
| 10 |
+
- **Model**: facebook/vitb-16
|
| 11 |
+
- **Dictionary Size**: 49152
|
| 12 |
+
- **Input Dimension**: 768
|
| 13 |
+
- **Expansion Factor**: 64
|
| 14 |
+
- **CLS Token Only**: False
|
| 15 |
+
|
| 16 |
+
### Training
|
| 17 |
+
- **Learning Rate**: 0.0061296854816475805
|
| 18 |
+
- **Batch Size**: 4096
|
| 19 |
+
- **Context Size**: 50
|
| 20 |
+
|
| 21 |
+
### Sparsity
|
| 22 |
+
- **L0 (Active Features)**: 661
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
## Additional Information
|
| 26 |
+
- **Wandb Run**: https://wandb.ai/perceptual-alignment/DINO-SAE-Vanilla-Sweep/runs/6bdgu49a/
|
| 27 |
+
- **Random Seed**: 42
|