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-2-2B
    4. Attention Out - 16k
    5. 0-GEMMASCOPE-ATT-16K
    6. 0
    Prev
    Next
    INDEX
    Explanations

    attends from sequences of dots to sequences of dots.

    oai_attention-head · gemini-1.5-flashTriggered by @dmontg

    attends to empty or punctuated segments from tokens that also contain placeholders or commentary

    oai_attention-head · gpt-4o-miniTriggered by @bot

    attends from arbitrary tokens to sequences of ten consecutive dots.

    oai_attention-head · gpt-4oTriggered by @nicholasraphaelgt

    attends to tokens within a list of typical courtroom/legal terminology, focusing on references to judicial and legal concepts, especially nearby mention of "the" following legal case citations.

    oai_attention-head · gpt-4.1-nanoTriggered by @liuliujiujiu05
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GEMMA-2-2B @ 0-gemmascope-att-16k
    Configuration
    google/gemma-scope-2b-pt-att/layer_0/width_16k/average_l0_104
    Prompts (Dashboard)
    36,864 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    Features
    16,384
    Data Type
    float32
    Hook Name
    blocks.0.attn.hook_z
    Hook Layer
    0
    Architecture
    jumprelu
    Context Size
    1,024
    Dataset
    monology/pile-uncopyrighted
    Activation Function
    relu
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Head Attr Weights
    0:0.70
    1:0.04
    2:0.01
    3:0.01
    4:0.12
    5:0.03
    6:0.02
    7:0.03
    Negative Logits
     
    -0.41
    <eos>
    -0.41
    ↵↵↵
    -0.34
    ↵
    -0.33
    ↵↵
    -0.31
     \
    -0.29
     "
    -0.28
     «
    -0.28
    mathrm
    -0.28
    ↵↵↵↵
    -0.27
    POSITIVE LOGITS
     Majefty
    0.72
     Efq
    0.69
     Monfieur
    0.68
     purpoſe
    0.68
     Reſ
    0.66
     myſelf
    0.65
     Theſe
    0.65
     Jefus
    0.62
     ſeveral
    0.62
     itſelf
    0.61
    Activations Density 0.005%

    No Known Activations