© 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 EvalsExportsinterp-engineNEWAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-27B-IT
    3. 34-GEMMASCOPE-2-TRANSCODER-262K
    4. 168001
    Prev
    Next
    INDEX
    Explanations

    **Analyzing Neuron Behavior**The neuron seems to be recognizing patterns related to structuring text with headings, subheadings, and list formats, especially within markdown or similar markup languages.Here's a breakdown of the observed patterns:1. **MAX_ACTIVATING_TOKENS**: This list prominently features terms like `headers`, `headings`, `and`, and numbers (`3`, `2`). The repetition of `headers` and `headings` strongly suggests a focus on structural elements of text.2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: This list contains symbols and words commonly found in text structuring and markup, such as `#` (often used for Markdown headings), `/`, `,`, `.`, `subtitles`, and `Include`. These tokens frequently follow structural markers.3. **TOP_POSITIVE_LOGITS**: This list includes terms that aren't immediately obvious in relation to text structure. However, when considering the context of generating content or instructions, some terms might indirectly relate to organization or specific types of output. For example, `Subdivision` could relate to breaking down content.4. **TOP_ACTIVATING_TEXTS**: These texts provide crucial context. They describe tasks like: * "guide of 1000 words optimized for seo with **headers h3, h3, lists** and a section FAQ" * "Always use **Markdown** with nice formatting to make it easier to follow. Organize it by **sections with headers**. Include references to the code as **markdown code blocks**" * "include **headings, subheadings, and lists**, aiming for a easy-to-read format." * "using **# headings # and # sub-headings #**" * "Provide **headings**. Include bullet point only in the **heading**" * "with **headings and subheadings** (including H2, H3 and H4 headings)" * "Use **sub headers** as header of each sub headers"**Synthesizing the Pattern**The overwhelming pattern across all lists, particularly the `MAX_ACTIVATING_TOKENS` and `TOP_ACTIVATING_TEXTS`, points to the neuron's activation by text that discusses or requires the use of **headings, subheadings, and structural formatting (like Markdown)** for organizing content. The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list reinforces this by showing the typical characters and words that follow such structural elements.**Concise Explanation:**markdown headings and subheadings

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-it/transcoder_all/layer_34_width_262k_l0_small_affine
    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
    ijoje
    0.46
    CTAssert
    0.43
    ാരി
    0.41
    SETS
    0.39
    tså
    0.38
    IED
    0.37
     McCulloch
    0.37
    czą
    0.37
    removeChild
    0.36
     aperiam
    0.36
    POSITIVE LOGITS
     Subdivision
    0.39
    wise
    0.38
    <'
    0.37
    த்திரம்
    0.36
     Late
    0.35
    llium
    0.35
     vigor
    0.35
    odor
    0.35
    bert
    0.34
     smoother
    0.34
    Activations Density 0.005%

    No Known Activations