© 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-2-27B
    3. 34-GEMMASCOPE-RES-131K
    4. 83702
    Prev
    Next
    INDEX
    Explanations

    thinksThe user wants me to explain the behavior of a neuron based on four lists: `TOP_POSITIVE_LOGITS`, `TOP_ACTIVATING_TEXTS`, `MAX_ACTIVATING_TOKENS`, and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`.The explanation should be concise (3-20 words) and capture what the neuron detects or predicts by finding patterns.I need to avoid specific phrases like "This neuron detects/predicts", "tokens", "patterns", and capitalization unless it's a proper noun.Let's analyze the provided lists:`MAX_ACTIVATING_TOKENS`: This list contains tokens that are highly activating for the neuron. I see symbols like ',', '+', '=', '-', '$', '(', '::*', '.', and numbers like '1'. It also includes words like 'value', 'added'.`TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: This list shows what comes *after* the `MAX_ACTIVATING_TOKENS`.- ',' is followed by 'n' or newlines.- '+' is followed by '?' or '2'.- '=' is followed by 'type' or '1'.- '-' is followed by '1'.- 'value' is followed by 'data'.- '(' is followed by 'n'.- 'added' is followed by a newline.- '::*' is followed by 'type'.- '.' is followed by 'apache'.`TOP_POSITIVE_LOGITS`: This list contains words/tokens strongly associated with this neuron. I see 'IERC', ';', 'visit', '()', 'besø', '))+', ']));'. These look like code snippets, especially related to programming languages (like C++, Java, or something similar). 'visit()' and 'besø' (Norse for visit) might suggest function calls or object methods. 'IERC' is an interface.`TOP_ACTIVATING_TEXTS`: This list gives context.- Contains examples like "Calculate the remainder when 16 is divided by g(-3", "Differentiate 2 + 6*l**3 + 7*l**3 + r". These are mathematical expressions.- "Header$type=menu", "World Conqueror 3 Apk Android Game". These suggest structured data or metadata, possibly related to web/app content.- "use strict"; "// Manually added data to be used by sbcs codec". This is JavaScript code.- "testConstructorMessageCause(org.apache.commons.math.FunctionEvaluationExceptionTest)". This is Java code, specifically mentioning a library (`org.apache.commons.math`).Combining these observations:- The `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` show a mix of mathematical operators, symbols, numbers, and code-related tokens.- `TOP_POSITIVE_LOGITS` points heavily towards code syntax, function calls, and interfaces.- `TOP_ACTIVATING_TEXTS` confirms this, showing mathematical formulas, code snippets (JavaScript, Java), and what looks like metadata descriptions or game titles with structured data. code-related symbols and keywords

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-27b-pt-res/layer_34/width_131k
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    naies
    -0.99
     خریدار
    -0.96
    🛻
    -0.96
     تأثير
    -0.95
     Suède
    -0.95
     Warenkorb
    -0.95
    お疲れ
    -0.94
    lione
    -0.93
     fonctionnent
    -0.93
    違います
    -0.93
    POSITIVE LOGITS
    IERC
    1.09
    ;
    
    1.01
     visit
    0.92
    ();
    
    0.92
     besø
    0.91
    ));
    
    0.90
    yka
    0.90
    UnityEngine
    0.90
    )}+\
    0.89
    ]));
    
    0.89
    Activations Density 0.011%

    No Known Activations