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. Residual Stream - 16k
    5. 0-GEMMASCOPE-RES-16K
    6. 343
    Prev
    Next
    INDEX
    Explanations

    words related to the concept of companionship or teamwork

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

    The token "iz" frequently appears in proper names, especially in non-English names or words. It often occurs in the middle or end of names for people, places, or technical terms.

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

    the names of people, likely authors or contributors, appearing in a list of references or citations.

    oai_token-act-pair · gemini-1.5-flashTriggered by @johnny

    The examples show proper nouns, often names of people or places, with a consistent substring appearing within them. These substrings might represent a common element in names from a certain region or cultural group, a shared etymological root, or an artifact of data collection.

    eleuther_acts_top20 · gemini-1.5-flashTriggered by @johnny

    In the text, specific character sequences like "iz", "iz>>", and "iz\n" are used to represent a variety of expressions such as names, titles, and abbreviations. These sequences typically appear in conjunction with words, acronyms, or phrases in contexts specific to the content of the text.

    eleuther_acts_top20 · gpt-3.5-turboTriggered by @johnny

    The token “iz” as part of names, primarily as a suffix of personal and geographical names across various cultures and languages.

    eleuther_acts_top20 · gpt-4-turboTriggered by @johnny
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GEMMA-2-2B @ 0-gemmascope-res-16k
    Configuration
    google/gemma-scope-2b-pt-res/layer_0/width_16k/average_l0_105
    Prompts (Dashboard)
    36,864 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    Features
    16,384
    Data Type
    float32
    Hook Name
    blocks.0.hook_resid_post
    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

    Negative Logits
     AssemblyVersion
    -0.72
     Normdatei
    -0.64
    adə
    -0.61
    ագրություններ
    -0.60
     Auger
    -0.57
    sile
    -0.55
     McMaster
    -0.55
     Govern
    -0.54
    gouv
    -0.54
     monographs
    -0.54
    POSITIVE LOGITS
     Choices
    0.83
     choices
    0.74
    choices
    0.71
    tagHelperRunner
    0.68
    choice
    0.68
    CHOICE
    0.67
     alternatives
    0.66
    iz
    0.65
     Alternatives
    0.65
    CHOICES
    0.63
    Activations Density 0.103%

    No Known Activations