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

    Neuronpedia

    APIAssistant AxisNEWCircuit TracerNEWSteerSAE EvalsExports Community BlogPrivacy & TermsContact
    1. Home
    3. Google DeepMind · Exploring Gemma 2 with Gemma Scope
    5. Gemma-2-9B-IT
    7. Residual Stream - 131k
    9. 20-GEMMASCOPE-RES-131K
    11. 113175
    Prev
    Next
    INDEX
    Explanations

    programming structures and conditional statements in code

    oai_token-act-pair · gpt-4o-miniTriggered by @bot

    code snippets or log messages (`');`, `dom.createDom`, `06;`)2. HTML/XML tags (`</a>`, `<h3>`, `<ul>`)3. part of code structure (`new FlowLayout`, `filterTF.getDocument().add`)4. structured text (`|-DT - Determiner`)5. command-line or control flow markers (`<start_of_turn>user`)The MAX_ACTIVATING_TOKENS (`>`, `"`, `-`, `(`, `<`) combined with TOKENS_AFTER_MAX_ACTIVATING_TOKEN suggest detecting delimiters, quotation marks, or structural markers within text/code. The `TOP_POSITIVE_LOGITS` are a bit of a red herring with their diverse languages, but "Charge" or "further" could potentially appear in logs or specific contexts.Given the overwhelming presence of code/structured text examples, and the punctuation in MAX_ACTIVATING_TOKENS, the neuron seems to be looking for delimiters and control characters common in programming or data formats.Final Answer: code delimiters

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GEMMA-2-9B-IT @ 20-gemmascope-res-131k
    Configuration
    google/gemma-scope-9b-it-res/layer_20/width_131k/average_l0_81
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    Features
    131,072
    Data Type
    float32
    Hook Name
    blocks.20.hook_resid_post
    Hook Layer
    20
    Architecture
    jumprelu
    Context Size
    1,024
    Dataset
    monology/pile-uncopyrighted
    Activation Function
    relu
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     betweenstory
    -0.75
    DoubleQuotes
    -0.74
    TypedDataSet
    -0.73
     autorytatywna
    -0.65
     AssemblyTitle
    -0.62
     دیکھیے
    -0.62
    Personendaten
    -0.61
    tagHelperRunner
    -0.60
     pinulongan
    -0.60
     CURIAM
    -0.60
    POSITIVE LOGITS
     nữa
    0.34
     further
    0.33
     en
    0.31
    đ
    0.31
     notably
    0.31
     Charge
    0.31
     dep
    0.30
    편
    0.30
     предо
    0.30
     charge
    0.30
    Activations Density 0.117%

    No Known Activations