Skip to content

ai-safety-foundation/sparse_autoencoder

Repository files navigation

Sparse Autoencoder

PyPI PyPI - License Checks Release

A sparse autoencoder for mechanistic interpretability research.

Read the Docs Here

Train a Sparse Autoencoder in colab, or install for your project:

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 , 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 for a guide to using this library.

Contributing

This project uses Poetry for dependency management, and PoeThePoet 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.

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).

poe