Introduction

This codebase is for "Sparse Autoencoders Can Interpret Randomly Initialized Transformers" and is heavily based on the following repositories:

• EleutherAI/sae
• EleutherAI/sae-auto-interp
• adamkarvonen/SAEBench

Installation

Required packages:

pip install torch
pip install transformers
pip install datasets
pip install nnsight
pip install numpy
pip install pandas
pip install matplotlib
pip install seaborn
pip install scikit-learn
pip install orjson

Optional dependencies:

pip install bitsandbytes  # For 8-bit model loading

Code Organization

The codebase is organized into several components:

• sae/: Core sparse autoencoder implementation
• sae-auto-interp/: Automated interpretation tools and utilities
• SAEBench/: Evaluation and benchmarking tools
• experiments/: Scripts for running the training, evaluation and analysis experiments

All experiment scripts are prefixed with "run_" in the experiments folder, with different configurations for various model sizes and settings.