Neuronpedia Export & Push

Neuronpedia is the community platform for browsing SAE features. miStudio can share your findings two ways:

1. Export to ZIP — build a complete, Neuronpedia-compatible package for manual upload or offline archival
2. Direct Push — push features, weights, and dashboard data straight into a local Neuronpedia instance over its API

Both run as background jobs with live progress, so multi-thousand-feature exports don't tie up your browser.

Export to ZIP

From an SAE card, choose Export and configure:

Option	Description
Feature selection	All features, extracted-only, or a custom index range
Logit lens data	Promoted/suppressed tokens per feature (top-K configurable)
Activation histograms	Distribution of each feature's activation values (bin count configurable)
Top activating tokens	Aggregated most-frequent tokens per feature
Explanations	Your feature labels, human and LLM-generated
SAELens weights	cfg.json + sae_weights.safetensors for Python interop

What's in the package

export.zip
├── metadata.json            # SAE config + export manifest
├── README.md                # human-readable summary
├── features/                # one JSON per feature
│   ├── 0.json               #   stats, logit lens, histogram,
│   ├── 1.json               #   top tokens, activation examples
│   └── ...
├── explanations/
│   └── explanations.json    # all feature labels
└── saelens/
    ├── cfg.json             # SAELens-compatible config
    └── sae_weights.safetensors

The saelens/ directory loads directly in SAELens/TransformerLens research code, so a single export serves both Neuronpedia upload and Python analysis.

Progress reports the current stage (computing logit lens, generating histograms, packaging) and the export can be cancelled mid-run. Completed exports stay in the export history for re-download until deleted.

Logit lens, briefly

For each feature, miStudio projects the feature's decoder direction through the model's unembedding matrix. Tokens with the highest projection are the ones the feature promotes when active; the most negative are suppressed. It's a fast, surprisingly informative summary of what a feature does to the model's output — see the Interpretability Primer for context.

Direct Push to Local Neuronpedia

If you run a local Neuronpedia instance (miStudio ships a K8s manifest for one), Push to Neuronpedia uploads everything in one step:

1. miStudio checks the instance is reachable (a status indicator shows this before you start)
2. A background job creates the model + SAE records in Neuronpedia with names derived from your SAE for discoverability
3. Dashboard data — logit lens, feature statistics, activation histograms — is computed on the fly and pushed per feature
4. Progress (features pushed / total) streams live; failures report the exact error

The push job survives browser refreshes — it runs server-side, and the panel re-attaches to its progress when you return.

Compute dashboard data separately

The Compute Dashboard Data action pre-computes logit lens + histograms for an SAE without pushing. Useful when you want the data cached before a large push, or for local inspection only.

Which Mode Should I Use?

Situation	Use
Sharing with the public community	Export ZIP → upload to neuronpedia.org
Lab-internal feature browsing	Direct push to your local instance
Archiving results / reproducibility	Export ZIP (self-contained)
Python analysis in SAELens	Export ZIP → use the saelens/ folder