Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    APICircuit TracerNEWSteerSAE EvalsExportsSlackBlogPrivacy & TermsContact
    © Neuronpedia 2025
    Privacy & TermsBlogGitHubSlackTwitterContact
    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. 75426
    Prev
    Next
    INDEX
    Explanations

    occurrences of recognizable entities, including names and places

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

    dates and years, particularly from the late 19th to early 21st centuries.

    oai_token-act-pair · claude-3-5-sonnet-20240620Triggered by @nrbln

    years between 1900 and 2099.

    oai_token-act-pair · gemini-1.5-proTriggered by @nrbln

    Dates and years are often marked as important, especially when referring to historical events or founding dates of organizations. The word "the" is frequently highlighted, particularly when preceding important nouns or concepts. Some examples also emphasize personal pronouns like "it" or "she" when referring to organizations or individuals.

    eleuther_acts_top20 · claude-3-5-sonnet-20240620Triggered by @nrbln

    The marked words often include references to years, centuries, ordinal numbers, and unfilled placeholders within numerical or formatted contexts, highlighting missing or specific numeric or positional content frequently tied to historical, chronological, or formal identification.

    eleuther_acts_top20 · gpt-4oTriggered by @nrbln

    names, dates, or objects

    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
    ActionCreators
    -0.38
     psychi
    -0.38
    fizierung
    -0.37
     utanför
    -0.37
     neutre
    -0.35
     fusil
    -0.35
     Autorisations
    -0.35
     Niedersachsen
    -0.35
     minutes
    -0.34
     charla
    -0.34
    POSITIVE LOGITS
    uxxxx
    0.68
    ArrowToggle
    0.61
     ModelExpression
    0.60
    EndContext
    0.56
     للاسماء
    0.52
    LookAnd
    0.51
    tanleria
    0.51
    thâu
    0.49
    AndEndTag
    0.48
    hyrchwyd
    0.47
    Activations Density 0.039%

    No Known Activations