Neuronpedia Is Open Source
All of Neuronpedia's code is open source, under the Apache 2.0 license.
GitHub: github.com/hijohnnylin/neuronpedia
You can run your own copy of Neuronpedia, for example to use your own private models and SAEs, or to make changes to it. Read why we made Neuronpedia open source.
What Is in the Repo
Neuronpedia is a group of apps. Each app runs by itself.
| App | What it does |
|---|---|
webapp | The website and the public API. Next.js and Postgres. |
inference | Runs the models for activations, steering, search, and the Jacobian Lens. |
autointerp | Makes and scores explanations. |
graph | Makes attribution graphs. |
nla | Runs Natural Language Autoencoders. |
sparsity | Runs weight-sparse transformers. |
The inference, graph, and nla apps use interp-engine to run models.
Run It Yourself
The README has the full steps. In short:
- Set up a database. Install Postgres 16+ with the pgvector extension, then run
make db-init. - Start the webapp. Run
make webapp-install, thenmake webapp-dev. Go to localhost:3000. - Import data. Your database is empty at the start. Use the admin panel to import sources from the exports.
- Start other apps that you need. For example, run
make inference-devfor steering and test activations. This needs a GPU for most models.
Run make help to see all the commands.
Use Your Own Data
To make dashboards for your own SAEs, see the README section "I Want to Generate My Own Dashboards/Data and Add It to Neuronpedia". You can also upload your SAEs to neuronpedia.org.
Contribute
We accept pull requests. Read the contributing section of the README first. You can also help with:
- Data - SAEs, attribution graphs, Jacobian lenses, and NLAs for new models.
- Bugs and ideas - Open an issue on GitHub, or tell us on Slack.