© 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. Google DeepMind · Exploring Gemma 2 with Gemma Scope
    3. Gemma-2-9B-IT
    4. Residual Stream - 131k
    5. 20-GEMMASCOPE-RES-131K
    6. 74618
    Prev
    Next
    INDEX
    Explanations

    dialogue and conversational exchanges

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

    thinksThe user wants me to explain the behavior of a neuron based on the provided lists.The explanation should be a concise phrase (3-20 words).I need to find a pattern related to tokens, text, or positive logits.Let's examine the lists:- **MAX_ACTIVATING_TOKENS**: These are mostly quotation marks (`"`, `“`, `‘`, `:`) and the word "that". This suggests the neuron might be related to direct speech or quoting something.- **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: These are words like "It", "I", "Please", "yes", "Ah", "A", "Indeed". These seem to be follow-ups or responses, often at the beginning of a sentence or thought.- **TOP_POSITIVE_LOGITS**: "للمعارف", "uska", "QMetaType", "Personensuche", "kaynağından", "مواليد", "GOTREF", "setVerticalGroup", "الحره", "LabelTagHelper". This list contains words in various languages, some technical terms, and some general terms. It's a bit diverse but hints at potentially structured or specific information.- **TOP_ACTIVATING_TEXTS**: This list contains examples of text. - "I gave it to him, and he said, `Where is that?', and I said, `It's at the Headquarters Building,' and he said..." - This is a dialogue with quoted speech. The `MAX_ACTIVATING_TOKENS` include quotes and "that". - "she admitted. 'And how did you come by this bit of information?' the Queen inquired further." - More dialogue. - "Please, your gracious Majesty, I beg you please don't make me… 'Nobleman<end_of_turn>" - Dialogue. - "ok <Ankman> where are you from? <qwebirc10526> I'm from france, but i live in Montreal since 3 weeks, and i'll stay permanently" - Chat log, informal. - “I ask them, ‘do you feel like a good sewer?’ and they say ‘yes.’” - Dialogue. - "Detective: Do you want to talk to me? Defendant: What about? Detective: Ah, what, ah ... you can talk to me, you can tell me to pound sand. You know, those" - Dialogue. - "'Is my Lady prepared to render her verdict in the matter of Lady Messalina?' the Conqueror asked. 'I am, my Lord,' the Queen replied." - Dialogue. - "makes those decisions on whether a case gets filed criminally?” The witness responded, “A complaint’s made to a police department or sheriff’s department and they make that decision in conjunction" - Dialogue, with quotes. - "'What have you got to say for yourselves?' the Conqueror's voice was ominous. 'Indeed not, but when her Majesty the Queen asked the question, Lady Messalina threw a glance at the two of you" - Dialogue. - "She replied that the defendant was the only one she could think of. dialogue quotes

    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
    lacable
    -0.50
    neſs
    -0.47
     jspb
    -0.45
    hagic
    -0.43
    apunov
    -0.43
    makeText
    -0.42
    WriteBarrier
    -0.42
    制
    -0.42
     contagio
    -0.42
     BrowserModule
    -0.41
    POSITIVE LOGITS
     للمعارف
    0.54
     uska
    0.48
    QMetaType
    0.44
    Personensuche
    0.43
     kaynağından
    0.42
     مواليد
    0.40
    GOTREF
    0.39
    setVerticalGroup
    0.38
     الحره
    0.37
    LabelTagHelper
    0.37
    Activations Density 0.037%

    No Known Activations