This repository contains implementations for mechanistic interpretability research, focusing on understanding the internal representations and circuits of transformer language models.
Implementation of sparse autoencoders for analyzing transformer activations. Features include:
- Training sparse autoencoders on MLP layer activations
- 8x expansion factor with L1 sparsity penalty
- Token-aligned activation extraction and analysis
- Neuron patching and intervention experiments
- Comprehensive training monitoring and visualization
See sparse-autoencoder/README.md for detailed documentation.
Safety-focused circuit discovery and analysis framework using library. Features include:
- Automated circuit pattern mining across model layers
- Safety benchmark evaluation (deception, harmful content, manipulation, power-seeking)
- Contrasting example analysis for safety feature identification
- Interactive dashboard for circuit visualization
- Graph-based circuit representation and analysis
See circuit-tracer-playground/README.md for detailed documentation.