© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    Jacobian LensNEW
    Natural Language
    Autoencoders
    NEW
    Assistant AxisNEWCircuit TracerUPDATESteerSAE EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-12B-IT
    3. 12-GEMMASCOPE-2-RES-16K
    4. 13525
    Prev
    Next
    INDEX
    Explanations

    1. **Analyze MAX_ACTIVATING_TOKENS**: This list contains symbols like `_`, `<`, `(...)`, and `...`. These are often used as separators, placeholders, or structural elements in text or code.2. **Analyze TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: This list includes `time`, `**`, `var`, `pt`, `style`, `char`, `: ` , `data`, `and`, and `` ` ``. Many of these are programming-related (var, data, char, style), time-related (time), or structural (**, : , `).3. **Analyze TOP_POSITIVE_LOGITS**: This list contains tokens from various languages (`い`, `sensibilities`, `ا`, `री`, `тона`, `ைகளில்`, `து`, `isées`, `হাম্ম`, `disciplines`). This suggests the neuron might not be tied to a single language but perhaps to a broader concept or structure that spans languages.4. **Analyze TOP_ACTIVATING_TEXTS**: * "If {{time\_or\_where}}, the {{who}} {{what}}." - This looks like a template, possibly for generating text or understanding grammatical structure. It uses placeholders. * "complexity of the prompt: **1. Cognitive Abilities & Proficiency (The Foundation):** * **"Lu-"** derives from Latin..." - This suggests structure, lists, or explanations of terms. The bolding `**` is prominent. * "hens (-) in variable names. * **Why other options are valid:** * **B: my\_var:** Uses an underscore, which is allowed. * **C: \_myvar:** Starts" - Here, underscores `_`, hyphens `-`, and variable name structure are discussed. This strongly points to code or programming concepts. * "of the passage, 'group': Appropriate Group of the passage}}]") }} form, no other form. Input: {{pt came with c/o of fever, vomiting, and left leg pain }}<end\_of\_turn>" - This includes placeholder-like syntax `{{...}}` and medical abbreviations (`pt`, `c/o`). * "HTML element with the `style` attribute)" - Mentions HTML and `style` attribute.This neuron is detecting patterns related to placeholders, code elements, and structuring text. Looking at the lists:- **MAX_ACTIVATING_TOKENS**: `_`, `<`, `(...)`, `...` - These are structural or placeholder symbols, often found in code or templates.- **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `var`, `data`, `style`, `char`, ``: `, `` ` `` - These are common in programming, data, or stylistic definitions.- **TOP_ACTIVATING_TEXTS**: Examples include template syntax `{{...}}`, code-like structures `my_var`, `_myvar`, mentions of `style` attributes, and medical abbreviations `pt`.The recurring theme is the use of specific symbols and keywords that define structure, placeholders, or code elements.Therefore, a good explanation would be:placeholders and code structure

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-12b-it/resid_post/layer_12_width_16k_l0_medium
    Prompts (Dashboard)
    238,145 prompts, 512 tokens each
    Dataset (Dashboard)
    lmsys + oasst1
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     Edition
    1.13
    ﻲ
    1.13
    ى
    1.09
     AOL
    1.03
    lık
    1.02
    ının
    1.02
    alq
    0.99
    al
    0.98
    ान
    0.98
     CERN
    0.98
    POSITIVE LOGITS
    い
    1.28
     sensibilities
    1.24
    ا
    1.14
    री
    1.13
    тона
    1.11
    ைகளில்
    1.10
    து
    1.09
    isées
    1.07
    হাম্ম
    1.07
     disciplines
    1.05
    Activations Density 0.051%

    No Known Activations