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. 20-GEMMASCOPE-RES-16K
    6. 7577
    Prev
    Next
    INDEX
    Explanations

    programming-related code elements and structures

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

    Code, forms, and programming-related terms

    np_acts-logits-general · gemini-2.0-flash

    form HTML"}, "related_to": "LoginForm, Form, FORM, frm, formulaire, Frm, Form"}This neuron seems to be related to form elements, particularly in programming contexts (like `LoginForm`, `Form`, `frm`, `endphp`, `DataSource`).It could also be related to programming code or data structures.Let's re-examine the lists considering this:- **TOP_POSITIVE_LOGITS**: `LoginForm`, `Form`, `FORM`, `frm`, `formulaire`, `endphp`, `AutoScaleMode`, `Frm`. This strongly points to forms, UI elements, or programming code related to forms.- **MAX_ACTIVATING_TOKENS**: `.`, `we`, `t`, `neo`, `in`. - The text `dataGridView1.DataSource = tabula; } A: That code should work.` contains `DataSource`. The token `.` precedes it. `DataSource` is very form/data related. - The text `scintilla1.Text = s + " " + s + " " + s + " " + s +` uses `sc` which could relate to scripting or editing components, often part of a GUI. - The text `neoantigens. We present the top 20 HLA alleles, top 20 genes, and detailed neoantigen`. The token `neo` appeared. `neoantigen` is a technical term. - The text `in bronchodilator efficacy.` The token `in` appeared. - The text `we nonetheless wish to note`. The token `we` appeared.- **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: - After `.`: `DataSource` -> related to forms/data. - After `we`: `sc` -> not clearly related to forms, maybe generic text. - After `t`: `/**` -> comment in code. form/code elements

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GEMMA-2-2B @ 20-gemmascope-res-16k
    Configuration
    google/gemma-scope-2b-pt-res/layer_20/width_16k/average_l0_71
    Prompts (Dashboard)
    36,864 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    Features
    16,384
    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
    utnik
    -0.46
     Belén
    -0.43
    <![
    -0.42
     Jerusalén
    -0.42
    ględ
    -0.41
    bcryptjs
    -0.40
     españolas
    -0.40
     MyBatis
    -0.40
     nahilalakip
    -0.40
    endenza
    -0.40
    POSITIVE LOGITS
     LoginForm
    0.81
     Form
    0.74
    Form
    0.65
    OGND
    0.58
     FORM
    0.58
     frm
    0.57
     formulaire
    0.57
    endphp
    0.57
    AutoScaleMode
    0.57
     Frm
    0.56
    Activations Density 0.129%

    No Known Activations