How psakthivel353/sae-feature-visualizer works
SAE Feature Visualizer is a research workflow for turning GPT-2 layer activations into sparse features, corpus-based interpretations, and interactive exploration. Users prepare a text corpus, collect hooked activations, train a Top-K sparse autoencoder, analyze feature activations and examples, then inspect results in the Streamlit dashboard; feature labels are an optional manual enrichment. The graph includes README-documented capabilities whose implementation was not fully excerpted, while omitting uncertain internal links where source evidence was unavailable.
Main entry points: dashboard.py Streamlit Dashboard · collect_activations.py Activation Collection · feature_analysis.py Feature Analysis · layer8_chunk_manifest.json Activation Cache
Components 16
Corpus and Activations
- Corpus Preparation
scripts/download_corpus.py - Text Corpus
data - Activation Collection
src/collect_activations.py - Forward Hooks
src/hooks.py - Activation Cache
cache/layer8_chunk_manifest.json
Sparse Modeling
- SAE Training
src/train_sae.py - Sparse Autoencoder
src/sae_model.py - Model Checkpoints
checkpoints
Feature Interpretation
- Feature Analysis
src/feature_analysis.py - Feature Statistics and Examples
cache/feature_stats.json - Manual Feature Labeling
scripts/label_features.py - Feature Labels
cache
Interactive Exploration
- Streamlit Dashboard
src/dashboard.py
Other components
- Researcher
- Corpus Providers
- GPT-2
Connections 23
- Corpus Providers → Corpus Preparation: provides text
scripts/download_corpus.py - Corpus Preparation → Text Corpus: writes corpus
scripts/download_corpus.py - Text Corpus → Activation Collection: supplies text
src/collect_activations.py - Activation Collection → GPT-2: runs inference
src/collect_activations.py - Activation Collection → Forward Hooks: attaches hooks
src/collect_activations.py - Forward Hooks → GPT-2: captures layer output
src/hooks.py - Activation Collection → Activation Cache: writes chunks
src/collect_activations.py - Activation Cache → SAE Training: supplies activations
src/train_sae.py - SAE Training → Sparse Autoencoder: trains model
src/train_sae.py - SAE Training → Model Checkpoints: saves weights
src/train_sae.py - Activation Cache → Feature Analysis: provides activations
src/feature_analysis.py - Model Checkpoints → Feature Analysis: loads model
src/feature_analysis.py - Feature Analysis → Sparse Autoencoder: encodes activations
src/feature_analysis.py - Feature Analysis → Feature Statistics and Examples: produces statistics and examples
src/feature_analysis.py - Feature Statistics and Examples → Manual Feature Labeling: guides labeling
scripts/label_features.py - Manual Feature Labeling → Feature Labels: saves labels
scripts/label_features.py - Researcher → Streamlit Dashboard: explores features
README.md - Feature Statistics and Examples → Streamlit Dashboard: displays results
src/dashboard.py - Model Checkpoints → Streamlit Dashboard: loads checkpoint
src/dashboard.py - Feature Labels → Streamlit Dashboard: shows labels
- Streamlit Dashboard → GPT-2: runs live inference
README.md - Streamlit Dashboard → Forward Hooks: uses activation hooks
src/dashboard.py - Streamlit Dashboard → Sparse Autoencoder: encodes live activations
src/dashboard.py