© 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. 16-GEMMASCOPE-2-RES-262K
    4. 106272
    Prev
    Next
    INDEX
    Explanations

    Code, section, lawsThinking Process:The user wants a concise explanation (3-20 words) of a neuron's behavior based on input lists.I need to find a pattern in `MAX_ACTIVATING_TOKENS`, `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`, `TOP_POSITIVE_LOGITS`, and `TOP_ACTIVATING_TEXTS`.1. **Analyze `MAX_ACTIVATING_TOKENS`**: `Code`, `{`, `zone`, `{`, `Code`, `Code`, `'`, `way`, `'`, `στην`. - "Code" appears multiple times. - `zone` might be a keyword or context. - `way` could be part of "two ways". - `στην` is Greek.2. **Analyze `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`**: `Section`, `.`, `print`, `Section`, `minds`, `,`, `when`, `ανα`. - "Section" appears multiple times. - "print" is common in coding. - `ανα` appears in a Greek phrase.3. **Analyze `TOP_POSITIVE_LOGITS`**: `b`, `p`, `g`, `س`, `sp`, `d`, `с`, `ok`, `y`, `for`. These are single letters or short strings, often seen in code (e.g., variables, keywords) or language identifiers. The Arabic letter 'س' and Russian letters 'с' suggest multilingual capabilities or potential representation of various symbols/characters in code or text.4. **Analyze `TOP_ACTIVATING_TEXTS`**: - "California Penal **Code** **Section** 632" - "`awk` **script**" - "end **zone**" - "**Penal** **Code** 21212" - "**Penal** **Code** **Section** 632" - "to-do list" - "`worldbuilding`" - "Cognitive Behavioural Therapy" - Greek text: "απλοποιημένη μέθοδο βασισμένη στην αναγωγή ψευ" (simplified method based on reduced pseudo [likely referring to programming/math concepts])**Pattern Identification**:- **Code/Programming**: "Code", "Code", "Code", "`awk` script", "`print`", "`worldbuilding`".- **Legal/Sections**: "Penal Code", "Penal Code Section", "Section", "zone" (as in end zone or legal zone).- **Multilingual/Symbols**: Greek tokens (`στην`, `ανα`, Greek text), Arabic/Russian characters in `TOP_POSITIVE_LOGITS`.- **Miscellaneous**: "way", "mind", "Therapy", "list".The most prominent and recurring themes that appear across `MAX_ACTIVATING_TOKENS`, `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`, and `TOP_ACTIVATING_TEXTS` are:- "Code" (from `MAX_ACTIVATING_TOKENS` and `TOP_ACTIVATING_TEXTS`)- "Section" (from `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` and `TOP_ACTIVATING_TEXTS`)- "laws" (implied by "Penal Code", "Penal Code Section" in `TOP_ACTIVATING_TEXTS`)- "zone" (from `MAX_ACTIVATING_TOKENS` and `TOP_ACTIVATING_TEXTS`)Considering the rule to be concise and specific, and not start with introductory phrases:"Code", "section", "laws", and "zone" all seem relevant.The phrase "code, section, laws" captures the legal and programming aspects strongly represented. "zone" is also present.Let's refine. The neuron seems to activate strongly on texts discussing legal codes, sections of law, and potentially programming code due to the presence of "Code" and "awk script".A combination like "legal code sections" or "code and laws" is good.The Greek parts suggest it might also generalize to other languages or technical terms, but "Code", "Section", and "Penal Code" are very strong signals."Code, Section, Laws" is 3 words, concise."Code, law sections, zone" is 4 words."Legal codes and programming" is 4 words.Looking at the `MAX_ACTIVATING_TOKENS`: `Code`, `zone`.Looking at `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `Section`, `print`.Looking at `TOP_ACTIVATING_TEXTS`: "Penal Code", "Section", "laws", "awk script", "worldbuilding", "end zone".The neuron frequently activates on legal documents or discussions involving specific "Code Sections" and also on programming contexts ("Code", "awk script"). The appearance of "zone" in both MAX_ACTIVATING_TOKENS and TOP_ACTIVATING_TEXTS is also notable.The combination of "Code" (both programming and legal) and "Section" (legal) seems to be the core. "Laws" is a good descriptor for the context. "zone" is also present.A concise phrase that combines these elements."Code, section, laws" covers the primary concepts well.Let's re-evaluate the possibility of including "zone". "End zone", "legal zone". It's present but perhaps less central than "Code" and "Section" across all snippets.The `TOP_POSITIVE_LOGITS` include `b`, `p`, `g`, `sp`, `d`, `y`, `for`. These could

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-it/resid_post/layer_16_width_262k_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
     dispose
    0.61
    توى
    0.57
     prodotto
    0.54
     relish
    0.53
     disulfide
    0.53
     harina
    0.52
     disseminate
    0.51
     déficit
    0.51
     disseminating
    0.51
     disposizione
    0.50
    POSITIVE LOGITS
    b
    0.77
    p
    0.72
    g
    0.72
    س
    0.69
    sp
    0.63
    d
    0.62
    с
    0.60
    ok
    0.59
    y
    0.57
    for
    0.56
    Activations Density 0.000%

    No Known Activations