Skip to main content

Exports

You can download Neuronpedia's data. This is faster than the API for a large number of features, and it does not use your rate limit.

Exports: neuronpedia-datasets.s3.us-east-1.amazonaws.com/index.html?prefix=v1/

The exports bucket, with one folder for each model.

Folders​

The exports use the same IDs as the website:

v1/
├── config/                    # explanation types, score types, and other settings
└── [MODEL_ID]/
    ├── [SOURCE_ID]/
    │   ├── features/          # batch-0.jsonl.gz, batch-1.jsonl.gz, ...
    │   ├── activations/       # top activations for each feature
    │   ├── explanations/      # auto-interp explanations
    │   ├── model.jsonl
    │   ├── release.jsonl
    │   ├── source.jsonl
    │   └── sourceset.jsonl
    └── headvis/[DATASET]/     # attention head data, in the HeadVis format

Each .jsonl.gz file is a gzipped file with one JSON object on each line.

Some sources have more than one explanations folder, for example explanations-gpt-4o-mini/. Each folder has the explanations from a different AI model.

Example - Download With the AWS CLI​

The bucket is public, so you do not need an AWS account:

aws s3 sync --no-sign-request \
  s3://neuronpedia-datasets/v1/gemma-2-2b/20-gemmascope-res-16k/explanations/ \
  ./explanations/

Load Exports Into Your Own Neuronpedia​

If you run Neuronpedia yourself, you can import exports into your local database from the admin panel.

Missing Something?​

If you cannot find the data that you need, email us. We can usually export it for you within a few days.

Older Exports​

Some old exports are at neuronpedia-exports.s3.amazonaws.com.