Skip to main content

Upload Your SAEs

There are two ways to put your SAEs (or transcoders, or other features) on Neuronpedia.

Option 1: Ask Us​

Fill out this form. It takes less than 5 minutes. We make the dashboards, explanations, and inference for you. We usually reply within a few days.

Use this option if you want your SAEs to be listed in the Releases menu, or to have inference (steering, search, and test activations).

Option 2: Upload With Python​

Use the Neuronpedia Python library to upload features yourself. You must have a Neuronpedia API key.

pip install neuronpedia

The steps are:

  1. Make sure the model exists. Go to neuronpedia.org/[MODEL_ID]. If it does not exist, make it with Model.new(...).
  2. Make a source set. For example, a source set named puppyscope-res-2k makes one source for each layer: 0-puppyscope-res-2k, 1-puppyscope-res-2k, and more.
  3. Upload features. For each source, call source.upload_batch(...) with your features. Each feature can have activations, explanations, and logits.

Only the user who made a source set can upload features to it.

See the example notebook: upload_features.ipynb.

Features that you upload are at their URLs right away. They are not listed on the home page or in the default dropdowns.

Make Dashboards With SAEDashboard​

To make dashboard data (top activations, logits, and density) for your SAEs, use SAEDashboard. It works with SAELens SAEs.