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

    Neuronpedia

    Natural Language
    Autoencoders
    NEW
    Assistant AxisNEWCircuit TracerUPDATESteerSAE EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-12B-IT
    3. 24-GEMMASCOPE-2-RES-16K
    4. 1848
    Prev
    Next
    INDEX
    Explanations

    This neuron seems to be associated with specific item/object descriptions and their properties or actions.Let's break down the lists to find a pattern:1. **MAX_ACTIVATING_TOKENS**: `classic`, `cylinder`, `linen`, `form`, `!`, `both`, `supplement`, `realizing`, `,`, `#`2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `,`, `pants`, `column`, `Let`, `interested`, `?`, `that`, `target`, `Use`3. **TOP_POSITIVE_LOGITS**: These look like potential word fragments or unlikely tokens, making them harder to directly interpret for common patterns.This neuron is activated by specific nouns and action-oriented verbs.Let's analyze the data provided:* **MAX_ACTIVATING_TOKENS**: `classic`, `cylinder`, `linen`, `form`, `!`, `both`, `supplement`, `realizing`, `,`, `#`* **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `,`, `pants`, `column`, `Let`, `interested`, `?`, `that`, `target`, `Use`* **TOP_POSITIVE_LOGITS**: `кты`, `dete`, `HMO`, `combate`, `ေါ`, `cleanse`, `croche`, `pessimistic`, `estabele`, `rearview`* **TOP_ACTIVATING_TEXTS**: * "independent, **classic**, and foreign films." (Followed by minor punctuation) * " **cylinder** .name = name" (Followed by `.name`) * " **linen** dress," " **linen** **pants** ," etc. (Followed by `pants` or `etc.`) * "word "**form** column A" (Followed by `column`) * "it **supplement** ?This neuron activates on nouns and their descriptive or functional attributes.Let's look at the patterns:* **MAX_ACTIVATING_TOKENS**: `classic`, `cylinder`, `linen`, `form`, `!`, `both`, `supplement`, `realizing`, `,`, `#`* **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `,`, `pants`, `column`, `Let`, `interested`, `?`, `that`, `target`, `Use`* **TOP_POSITIVE_LOGITS**: `кты`, `dete`, `HMO`, `combate`, `ေါ`, `cleanse`, `croche`, `pessimistic`, `estabele`, `rearview`* **TOP_ACTIVATING_TEXTS**: * "...independent, **classic**, and foreign films." * " **cylinder** .name = name" * " **linen** dress," " **linen** **pants** ," etc. * "word "**form** column A" * "you **supplement** ?" * "you **realizing** that"**Analysis**:The `MAX_ACTIVATING_TOKENS` list contains specific nouns like `classic`, `cylinder`, `linen`, `form`, and `supplement`.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list shows tokens that often follow these nouns, such as `pants`, `column`, `interested`, `target`, `Use`. This suggests a relationship between an object/concept and its related properties, types, or actions.The `TOP_ACTIVATING_TEXTS` confirm this:* "classic" films* "cylinder" object/variable name* "linen" clothing items (dress, pants)* "form" column* "supplement" to a question* "realizing" a conceptThe `TOP_POSITIVE_LOGITS` are less clear and might relate to some very specific or low-level representations, or potentially non-English words. However, the pattern from the other lists is strong.The core behavior seems to be about identifying specific *things* or *concepts* (`classic`, `cylinder`, `linen`, `form`, `supplement`) and their immediate context, which often involves their type, a related object, or a question about them.**Concise Explanation**:specific nouns and their related items

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-12b-it/resid_post/layer_24_width_16k_l0_medium
    Prompts (Dashboard)
    238,145 prompts, 512 tokens each
    Dataset (Dashboard)
    lmsys + oasst1
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    Plas
    0.83
    IR
    0.78
    Play
    0.78
    Dial
    0.77
    Band
    0.75
    Aspect
    0.75
    Dar
    0.75
    BAND
    0.75
    Rus
    0.74
    Duty
    0.74
    POSITIVE LOGITS
    кты
    0.84
     dete
    0.84
     HMO
    0.82
     combate
    0.81
    ေါ
    0.78
     cleanse
    0.78
     croche
    0.77
     pessimistic
    0.77
     estabele
    0.77
     rearview
    0.77
    Activations Density 0.001%

    No Known Activations