Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    Neuronpedia is an open interpretability platform.
    Explore, steer, and experiment on AI models.
    Tutorial: Gemma Scope
    MIT Technology ReviewGoogle DeepMindOpenMOSS, Fudan UniversityApollo ResearchMATSEleutherAI
    Explore
    Browse over four terabytes of activations, explanations, and metadata. Neuronpedia supports probes, latents/features, custom vectors, concepts, and more.

    Releases

    Llama Scope R1: SAEs for DeepSeek-R1-Distill-Llama-8B
    OpenMOSS Team, Fudan University
    Gemma Scope - Exploring the Inner Workings of Gemma 2
    Language Model Interpretability Team, Google DeepMind
    AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders
    pyvene.ai, The Stanford NLP Group
    Llama Scope: SAEs for Llama-3.1-8B
    OpenMOSS Team, Fudan University
    Feature Splitting for GPT2-Small
    Joseph Bloom
    Gated SAE for Llama3-8B-Instruct
    Julius Han
    Multi TopK SAE for Llama3.1-8B
    EleutherAI
    Sparse Autoencoder for GPT2-Small - v5
    OpenAI
    Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning
    Apollo Research · Jordan Taylor
    Transcoders Enable Fine-Grained Interpretable Circuit Analysis for Language Models
    Jacob Dunefsky · Philippe Chlenski
    Sparse Autoencoders for Pythia-70M-Deduped
    Under Peer Review
    Attention SAE Research Paper
    Under Peer Review
    Open Source Sparse Autoencoders for all Residual Stream Layers of GPT2-Small
    Joseph Bloom

    Models

    DEEPSEEK-R1-LLAMA-8B
    DeepSeek-R1-Dist-Llama-8BDeepSeek
    GEMMA-2-2B-IT
    Gemma-2-2B-ITGoogle Deepmind
    LLAMA3-8B-IT
    Llama3-8B-ITMeta
    LLAMA3.1-8B
    Llama3.1-8BMeta
    GEMMA-2-2B
    Gemma-2-2BGoogle Deepmind
    GEMMA-2-9B-IT
    Gemma-2-9B-ITGoogle Deepmind
    GEMMA-2-9B
    Gemma-2-9BGoogle Deepmind
    P70M-D
    Pythia-70M-DedupedEleutherAI
    GPT2-SM
    GPT2-SmallOpenAI

    Jump To

    Jump to Source/SAE
    Jump to Feature
    INDEX
    Jump to Random
    Steer
    Modify model behavior by steering its activations using latents or custom vectors. Steering supports instruct (chat) and reasoning models, and has fully customizable temperature, strength, seed, etc.
    Steering example with a cat feature
    Search
    Search over 50,000,000 latents/vectors, either by semantic similarity to explanation text, or by running custom text via inference through a model to find top matches.

    Search via Inference

    Run Example Search
    API + Libraries
    Neuronpedia hosts the world's first interpretability API (March 2024) - and all functionality is available by API or Python/TypeScript libraries. Most endpoints have an OpenAPI spec and interactive docs.
    Steering example with a cat feature
    Inspect
    Go in depth on each probe/latent/feature with top activations, top logits, activation density, and live inference testing. All dashboards have unique links, can be compiled into sharable lists, and supports IFrame embedding, as demonstrated here.
    Who We Are
    Neuronpedia was created by Johnny Lin, an ex-Apple engineer who previously founded a privacy startup. Neuronpedia is supported by Decode Research, the Long Term Future Fund, and AISTOF.
    Get Involved
    Citation
    @misc{neuronpedia,
    title = {Neuronpedia: Interactive Reference and Tooling for Analyzing Neural Networks},
    year = {2023},
    note = {Software available from neuronpedia.org},
    url = {https://www.neuronpedia.org},
    author = {Lin, Johnny}
}