© 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. 17231
    Prev
    Next
    INDEX
    Explanations

    The neuron seems to be focusing on how words start or follow specific single letters, especially when those beginnings lead to descriptive or comparative phrases.Let's break this down:1. **MAX_ACTIVATING_TOKENS**: It's just the letter 'a' repeated many times. This indicates the neuron is heavily influenced when the input token is 'a'.2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: This list shows what comes *after* the 'a'. * naughty * ch (could be part of 'change', 'chameleon' etc.) * second * ceiling * guard * hamburger * hydr (could be part of 'hydration', 'hydrating') * restaurant * slap Many of these are nouns or adjectives starting with specific letter combinations. The 'a' acts as a prefix or starting point. What's interesting is the nature of these words: 'naughty', 'second', 'ceiling', 'guard', 'hamburger', 'restaurant', 'slap'.3. **TOP_POSITIVE_LOGITS**: This tells us what the neuron *contributes to* positively. * Non-English characters (difficult to interpret without knowing the script, but they might represent words or concepts). * 'acson' * 'oven' * 'minded' * '-minded' (suffix for compound words like 'open-minded', 'kind-minded') * 'myriad' The presence of 'minded' and '-minded' is a strong signal.4. **TOP_ACTIVATING_TEXTS**: This is where we see the context. * "...swallows like a **naughty doll**..." * "...quick to adapt to the fast fashion trends

    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
    olan
    -0.07
    å¾·åĭĴ
    -0.06
    produksi
    -0.06
    ara
    -0.06
     liken
    -0.06
    ément
    -0.06
    öll
    -0.06
    è¯ģåΏèĤ¡ä»½æľīéĻIJåħ¬åı¸
    -0.06
     ÙĦÛĮÙĨ
    -0.06
    veill
    -0.06
    POSITIVE LOGITS
    виÑĨа
    0.06
    ä¸ĢæĹ¦
    0.05
    éļıçĿĢ
    0.05
    acson
    0.05
    oven
    0.05
     minded
    0.05
    æ½ľ
    0.05
    -minded
    0.05
     myriad
    0.05
    ä¸ĢæĹ¦è¢«
    0.05
    Activations Density 0.011%

    No Known Activations