Language-Model-SAEs

This repo aims to provide a general codebase for conducting dictionary-learning-based mechanistic interpretability research on Language Models (LMs). It powers a configurable pipeline for training and evaluating Sparse Autoencoders and their variants, and provides a set of tools (mainly a React-based webpage) for analyzing and visualizing the learned dictionaries.

The design of the pipeline (including the configuration and some training detail) is highly inspired by the mats_sae_training project (now known as SAELens) and heavily relies on the TransformerLens library. We thank the authors for their great work.

News

• 2024.10.29 We introduce Llama Scope, our first contribution to the open-source Sparse Autoencoder ecosystem. Stay tuned! Link: Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders

• 2024.10.9 Transformers and Mambas are mechanistically similar in both feature and circuit level. Can we follow this line and find universal motifs and fundamental differences between language model architectures? Link: Towards Universality: Studying Mechanistic Similarity Across Language Model Architectures

• 2024.5.22 We propose hierarchical tracing, a promising method to scale up sparse feature circuit analysis to industrial size language models! Link: Automatically Identifying Local and Global Circuits with Linear Computation Graphs

• 2024.2.19 Our first attempt on SAE-based circuit analysis for Othello-GPT and found an example of Attention Superposition in the wild! Link: Dictionary learning improves patch-free circuit discovery in mechanistic interpretability: A case study on othello-gpt.

Installation

Currently, the codebase use pdm to manage the dependencies, which is an alternative to poetry. To install the required packages, just install pdm, and run the following command:

pdm install

This will install all the required packages for the core codebase. Note that if you're in a conda environment, pdm will directly take the current environment as the virtual environment for current project, and remove all the packages that are not in the pyproject.toml file. So make sure to create a new conda environment (or just deactivate conda, this will use virtualenv by default) before running the above command. A forked version of TransformerLens is also included in the dependencies to provide the necessary tools for analyzing features.

If you want to use the visualization tools, you also need to install the required packages for the frontend, which uses bun for dependency management. Follow the instructions on the website to install it, and then run the following command:

cd ui
bun install

bun is not well-supported on Windows, so you may need to use WSL or other Linux-based solutions to run the frontend, or consider using a different package manager, such as pnpm or yarn.

Launch an Experiment

We provide both a programmatic and a configuration-based way to launch an experiment. The configuration-based way is more flexible and recommended for most users. You can find the configuration files in the examples/configuration directory, and modify them to fit your needs. The programmatic way is more suitable for advanced users who want to customize the training process, and you can find the example scripts in the examples/programmatic directory.

To simply begin a training process, you can run the following command:

lm-saes train examples/configuration/train.toml

which will start the training process using the configuration file examples/configuration/train.toml.

To analyze a trained dictionary, you can run the following command:

lm-saes analyze examples/configuration/analyze.toml --sae <path_to_sae_model>

which will start the analysis process using the configuration file examples/configuration/analyze.toml. The analysis process requires a trained SAE model, which can be obtained from the training process. You may need launch a MongoDB server to store the analysis results, and you can modify the MongoDB settings in the configuration file.

Generally, our configuration-based pipeline uses outer layer settings as default of the inner layer settings. This is beneficial for easily building deeply nested configurations, where sub-configurations can be reused (such as device and dtype settings). More detail will be provided future.

Visualizing the Learned Dictionary

The analysis results will be saved using MongoDB, and you can use the provided visualization tools to visualize the learned dictionary. First, start the FastAPI server by running the following command:

uvicorn server.app:app --port 24577 --env-file server/.env

Then, copy the ui/.env.example file to ui/.env and modify the VITE_BACKEND_URL to fit your server settings (by default, it's http://localhost:24577), and start the frontend by running the following command:

cd ui
bun dev --port 24576

That's it! You can now go to http://localhost:24576 to visualize the learned dictionary and its features.

Development

We highly welcome contributions to this project. If you have any questions or suggestions, feel free to open an issue or a pull request. We are looking forward to hearing from you!

TODO: Add development guidelines

Citation

Please cite this library as:

@misc{Ge2024OpenMossSAEs,
    title  = {OpenMoss Language Model Sparse Autoencoders},
    author = {Xuyang Ge, Fukang Zhu, Junxuan Wang, Wentao Shu, Lingjie Chen, Zhengfu He},
    url    = {https://github.com/OpenMOSS/Language-Model-SAEs},
    year   = {2024}
}