DAG Model for gemmascope SAE
This repository contains a trained Directed Acyclic Graph (DAG) model for measuring effective L0 of a Sparse Autoencoder.
Model Info
- SAE Type: gemmascope
- SAE Release: gemma-scope-2b-pt-res
- SAE ID: layer_12/width_16k/average_l0_41
- d_sae: 16384
- Tokens Used: 10,000,000
- Effective L0: 21
- Actual L0: 47.7
- Compression Ratio: 2.27x
Files
final_model.safetensors: Trained DAG model (Lambda matrix, b_penalty, feature_order)results.json: Training metadata and metricstraining_curves.png: Loss curves and training progress visualization
Usage
Use with the Probabilistic SAE Streamlit dashboard:
- Check "Load pre-trained DAG from HF"
- DAG model HF repo:
TheodoreEhrenborg/dag-gemmascope-layer12-njotlwvk
- DAG model subfolder: (leave empty)
The dashboard will automatically load the matching SAE and enable clustering.
Training Details
Trained using effective_l0_vanilla.py with:
- Epochs: 5
- Learning rate: 0.0005
- Batch size: 6400
For more details, see results.json.
