© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    Jacobian LensNEW
    Natural Language
    Autoencoders
    NEW
    Assistant AxisNEWCircuit TracerUPDATESteerSAE EvalsExportsinterp-engineNEWAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Qwen3.5-4B
    3. 15-RES-MATRYOSHKA-65K
    4. 17980
    Prev
    Next
    INDEX
    Explanations

    The pattern I observe is related to observations or impressions, often followed by a descriptor or consequence.* "looks" followed by "amazing"* "alarming" followed by "circumstances"* "I" followed by "look" (in relation to appearance)* "What" followed by "is" (an exclamation/question about a state)* "someone" followed by "else" (indicating a different person)Considering "looks amazing," "alarming circumstances," "look 10 years younger," "What is it," and "someone else" – these suggest a focus on descriptions, states, or identification, sometimes with an emotional valence (alarming).The `TOP_POSITIVE_LOGITS` seem like noisy, potentially corrupted tokens, so I will rely more on the other three lists. The `TOP_ACTIVATING_TEXTS` contain phrases like "looks amazing," "alarming circumstances," "looks fantastic," "What an actress! What a faker!" "confused look," "What is it," "someone else."This neuron seems to capture descriptive statements or questions about appearance, state, or identity.Let's try to synthesize this into a concise phrase."appearance or state description" - 4 words."describing appearance or state" - 4 words."state or appearance descriptions" - 4 words."appearance and state observations" - 4 words."identifying appearance or states" - 4 words."states and appearances" - 3 words.states and appearances

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    decoderesearch/qwen-3.5-saes/qwen-3.5-4b
    Prompts (Dashboard)
    16,384 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    epass
    -0.06
     panggil
    -0.06
    ikia
    -0.06
    è¨Ī
    -0.05
     respiro
    -0.05
    -counter
    -0.05
    agée
    -0.05
    _gener
    -0.05
    åı¬éĽĨ
    -0.05
     aggreg
    -0.05
    POSITIVE LOGITS
    æĭīåħĭ
    0.06
    ç¡®
    0.06
    çİ°åľº
    0.05
    esor
    0.05
     thá»į
    0.05
    碾
    0.05
    azos
    0.05
    az
    0.05
    è°ģèĥ½
    0.05
    dur
    0.05
    Activations Density 0.022%

    No Known Activations