© 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. 31-GEMMASCOPE-2-RES-16K
    4. 15075
    Prev
    Next
    INDEX
    Explanations

    Analyzing Neuron Behavior**I'm currently focused on synthesizing the input lists to articulate a neuron's predictive function. My aim is to distill the core pattern it identifies, expressing this concisely without enumerating specific tokens.**Analysis:**1. **MAX_ACTIVATING_TOKENS**: The tokens are overwhelmingly `and` and `or`.2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: The tokens following `and/or` include words like `against`, `off`, `negative`, `down`, `who`, `y`, `after`. These suggest a relationship between conjunctions and contrasting elements, directional terms, or question words.3. **TOP_POSITIVE_LOGITS**: These are primarily non-English words, many appearing to relate to "incredible" or concepts often associated with positive sentiment or superlatives (e.g., the German "unglaublich" or Russian "невероят").**Pattern Identification:**The neuron is most activated by conjunctions (`and`, `or`). The words following these conjunctions (e.g., `against`, `off`, `negative`, `down`) suggest it might be identifying contrasts, negations, or divergent paths introduced by these conjunctions. The positive logits, being foreign words often meaning "incredible" or similar, hint at the neuron associating these conjunctions (especially when followed by specific types of words) with strong, perhaps surprising or emphatic statements.The texts show phrases like "arguments for **and** against", "switched on **or** off", "minimum **and** maximum", "scope *is* **and** *is not* included", "positive **or** negative", "scale your Redshift cluster up **or** down", "who is good **and** who is evil", "x **and** y coordinates", "before **and** after".This strongly suggests the neuron is detecting pairs of contrasting or related items connected by "and" or "or". The "positive logits" might be a byproduct of these contrasts being presented in a way that seems emphatic or "incredible".The core pattern is the use of conjunctions to link opposing or paired concepts.**Explanation:**"and/or" contrast pairs

    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_31_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
     others
    0.74
     other
    0.70
    other
    0.70
     ಇತರ
    0.66
    others
    0.65
     altri
    0.65
     інші
    0.65
    Other
    0.64
    maybe
    0.63
     інших
    0.62
    POSITIVE LOGITS
    🈚
    0.76
     بسیار
    0.67
    📟
    0.66
    เภท
    0.63
    ⛎
    0.62
    🏤
    0.61
     невероят
    0.60
    🕍
    0.60
    📠
    0.59
     unglaublich
    0.59
    Activations Density 0.105%

    No Known Activations