Note - I'm still open to accepting PRs on this library, and am very happy for other people to build on it, but I won't be actively maintaining it going forwards since I'll be focusing on my job. The SAELens library will continue to have more development and iteration, and it uses a fork of this repo as well as containing a much larger suite of tools for working with SAEs, so depending on your use case you might find that library preferable!

This codebase was designed to replicate Anthropic's sparse autoencoder visualisations, which you can see here. The codebase provides 2 different views: a feature-centric view (which is like the one in the link, i.e. we look at one particular feature and see things like which tokens fire strongest on that feature) and a prompt-centric view (where we look at once particular prompt and see which features fire strongest on that prompt according to a variety of different metrics).

Install with pip install sae-vis. Link to PyPI page here.

Important note - this repo was significantly restructured in March 2024 (we'll remove this message at the end of April). The recent changes include:

• The ability to view multiple features on the same page (rather than just one feature at a time)
• D3-backed visualisations (which can do things like add lines to histograms as you hover over tokens)
• More freedom to customize exactly what the visualisation looks like (we provide full cutomizability, rather than just being able to change certain parameters)

Here is a link to a Google Drive folder containing 3 files:

• User Guide, which covers the basics of how to use the repo (the core essentials haven't changed much from the previous version, but there are significantly more features)
• Dev Guide, which we recommend for anyone who wants to understand how the repo works (and make edits to it)
• Demo, which is a Colab notebook that gives a few examples

In the demo Colab, we show the two different types of vis which are supported by this library:

1. Feature-centric vis, where you look at a single feature and see e.g. which sequences in a large dataset this feature fires strongest on.

2. Prompt-centric vis, where you input a custom prompt and see which features score highest on that prompt, according to a variety of possible metrics.

To cite this work, you can use this bibtex citation:

@misc{sae_vis,
    title  = {{SAE Visualizer}},
    author = {Callum McDougall},
    howpublished    = {\url{https://github.com/callummcdougall/sae_vis}},
    year   = {2024}
}

This project is uses Poetry for dependency management. After cloning the repo, install dependencies with poetry install.

This project uses Ruff for formatting and linting, Pyright for type-checking, and Pytest for tests. If you submit a PR, make sure that your code passes all checks. You can run all checks with make check-all.

• 0.2.9 - added table for pairwise feature correlations (not just encoder-B correlations)
• 0.2.10 - fix some anomalous characters
• 0.2.11 - update PyPI with longer description
• 0.2.12 - fix height parameter of config, add videos to PyPI description
• 0.2.13 - add to dependencies, and fix SAELens section
• 0.2.14 - fix mistake in dependencies
• 0.2.15 - refactor to support eventual scatterplot-based feature browser, fix ’ HTML
• 0.2.16 - allow disabling buffer in feature generation, fix demo notebook, fix sae-lens compatibility & type checking
• 0.2.17 - use main branch of sae-lens
• 0.2.18 - remove circular dependency with sae-lens
• 0.2.19 - formatting, error-checking
• 0.2.20 - fix bugs, remove use of batch_size in config
• 0.2.21 - formatting