saev - Sparse Auto-Encoders for Vision

Implementation of sparse autoencoders (SAEs) for vision transformers (ViTs) in PyTorch.

This is the codebase used for our preprint "Sparse Autoencoders for Scientifically Rigorous Interpretation of Vision Models"

• arXiv preprint
• Huggingface Models
• API Docs
• Demos

About

saev is a package for training sparse autoencoders (SAEs) on vision transformers (ViTs) in PyTorch. It also includes an interactive webapp for looking through a trained SAE's features.

Originally forked from HugoFry who forked it from Joseph Bloom.

Read logbook.md for a detailed log of my thought process.

See related-work.md for a list of works training SAEs on vision models. Please open an issue or a PR if there is missing work.

Installation

Installation is supported with uv. saev will likely work with pure pip, conda, etc. but I will not formally support it.

To install, clone this repository (maybe fork it first if you want).

In the project root directory, run uv run python -m saev --help. The first invocation should create a virtual environment and show a help message.

Using saev

See the docs for an overview.

I recommend using the llms.txt file as a way to use any LLM provider to ask questions. For example, you can run curl https://osu-nlp-group.github.io/SAE-V/llms.txt | pbcopy on macOS to copy the text, then paste it into https://claude.ai and ask any question you have.