https://github.com/ai-safety-foundation/sparse_autoencoder

Sparse Autoencoder for Mechanistic Interpretability
https://github.com/ai-safety-foundation/sparse_autoencoder

Last synced: 10 days ago
JSON representation

Sparse Autoencoder for Mechanistic Interpretability

• Host: GitHub
• URL: https://github.com/ai-safety-foundation/sparse_autoencoder
• Owner: ai-safety-foundation
• License: mit
• Created: 2023-10-27T07:37:15.000Z (about 1 year ago)
• Default Branch: main
• Last Pushed: 2024-07-20T15:42:39.000Z (4 months ago)
• Last Synced: 2024-10-04T21:17:12.381Z (about 1 month ago)
• Language: Python
• Homepage: https://ai-safety-foundation.github.io/sparse_autoencoder/
• Size: 12 MB
• Stars: 175
• Watchers: 4
• Forks: 38
• Open Issues: 14
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

• awesome-llm-interpretability - Sparse Autoencoder - Sparse Autoencoder for Mechanistic Interpretability. (Table of Contents / LLM Interpretability Tools)

README

        
# Sparse Autoencoder

[![PyPI](https://img.shields.io/pypi/v/sparse_autoencoder?color=blue)](https://pypi.org/project/transformer-lens/)

![PyPI -

License](https://img.shields.io/pypi/l/sparse_autoencoder?color=blue) [![Checks](https://github.com/alan-cooney/sparse_autoencoder/actions/workflows/checks.yml/badge.svg)](https://github.com/alan-cooney/sparse_autoencoder/actions/workflows/checks.yml)

[![Release](https://github.com/alan-cooney/sparse_autoencoder/actions/workflows/release.yml/badge.svg)](https://github.com/alan-cooney/sparse_autoencoder/actions/workflows/release.yml)

A sparse autoencoder for mechanistic interpretability research.

[![Read the Docs

Here](https://img.shields.io/badge/-Read%20the%20Docs%20Here-blue?style=for-the-badge&logo=Read-the-Docs&logoColor=white&link=https://ai-safety-foundation.github.io/sparse_autoencoder/)](https://ai-safety-foundation.github.io/sparse_autoencoder/)

Train a Sparse Autoencoder [in colab](https://colab.research.google.com/github/ai-safety-foundation/sparse_autoencoder/blob/main/docs/content/demo.ipynb), or install for your project:

```shell

pip install sparse_autoencoder

```

## Features

This library contains:

   1. **A sparse autoencoder model**, along with all the underlying PyTorch components you need to

      customise and/or build your own:

      - Encoder, constrained unit norm decoder and tied bias PyTorch modules in `autoencoder`.

      - L1 and L2 loss modules in `loss`.

      - Adam module with helper method to reset state in `optimizer`.

   2. **Activations data generator** using TransformerLens, with the underlying steps in case you

      want to customise the approach:

      - Activation store options (in-memory or on disk) in `activation_store`.

      - Hook to get the activations from TransformerLens in an efficient way in `source_model`.

      - Source dataset (i.e. prompts to generate these activations) utils in `source_data`, that

        stream data from HuggingFace and pre-process (tokenize & shuffle).

   3. **Activation resampler** to help reduce the number of dead neurons.

   4. **Metrics** that log at various stages of training (e.g. during training, resampling and

      validation), and integrate with wandb.

   5. **Training pipeline** that combines everything together, allowing you to run hyperparameter

      sweeps and view progress on wandb.

## Designed for Research

The library is designed to be modular. By default it takes the approach from [Towards

Monosemanticity: Decomposing Language Models With Dictionary Learning

](https://transformer-circuits.pub/2023/monosemantic-features/index.html), so you can pip install

the library and get started quickly. Then when you need to customise something, you can just extend

the class for that component (e.g. you can extend `SparseAutoencoder` if you want to customise the

model, and then drop it back into the training pipeline. Every component is fully documented, so

it's nice and easy to do this.

## Demo

Check out the demo notebook [docs/content/demo.ipynb](https://github.com/ai-safety-foundation/sparse_autoencoder/blob/main/docs/content/demo.ipynb) for a guide to using this library.

## Contributing

This project uses [Poetry](https://python-poetry.org) for dependency management, and

[PoeThePoet](https://poethepoet.natn.io/installation.html) for scripts. After checking out the repo,

we recommend setting poetry's config to create the `.venv` in the root directory (note this is a

global setting) and then installing with the dev and demos dependencies.

```shell

poetry config virtualenvs.in-project true

poetry install --with dev,demos

```

### Checks

For a full list of available commands (e.g. `test` or `typecheck`), run this in your terminal

(assumes the venv is active already).

```shell

poe

```