© 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-2-27B
    3. 22-GEMMASCOPE-RES-131K
    4. 13956
    Prev
    Next
    INDEX
    Explanations

    path. `Sl` * `www.slicer.org/copyright/copyright.txt`: Contains `www`. The token after `www` would be `.`. The provided list has `%`. Again, generalization.Let's consolidate observations:* **MAX_ACTIVATING_TOKENS**: `Sl`, `path`, `2`, `h`, `www`.* **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `ns` (after `Sl`), `%`, `.`, `0`, `\`, `"`. `Sl` is often followed by `ns` (e.g., "Slns"). This looks like it's detecting specific suffixes or word parts.`path` is often a precursor to web addresses or file paths.`www` is explicitly a web prefix.Looking at `TOP_POSITIVE_LOGITS`: `ArgumentParser`, `CodeDom`, `collega`, `ersche`, `barten`. These are either technical terms or specific word components. `प्रोड`, `স্ট`, `инструменты` are foreign language words.The neuron seems to be triggered by certain word beginnings or specific tokens that might lead to technical terms, foreign words, or specific abbreviations/suffixes.The most striking pattern in `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` is `Sl` followed by `ns`. This forms a distinct unit often seen in German words or technical compound words.Let's examine the `TOP_POSITIVE_LOGITS` again:`collega` -> common in Latin-based languages for colleague.`ersche` -> part of common German words like "erscheinen".`barten` -> related to "Barten" (whales) or "arten" (types) in German.`invocations` -> not in logits, but `invoca` could suggest patterns.The foreign language words (`प्रोड`, `স্ট`, `инструменты`) and the partially translated/foreign words (`collega`, `ersche`, `barten`) suggest a focus on linguistic diversity or specific roots.Given `Sl` followed by `ns` and the presence of foreign and technical words, the neuron might be detecting specific linguistic patterns or origins. The foreign words in `TOP_POSITIVE_LOGITS` strongly suggest looking for non-English elements.Let's re-evaluate `MAX_ACTIVATING_TOKENS`: `Sl`, `path`, `2`, `h`, `www`.`Sl` -> `ns``path` -> `%``2` -> `.``h` -> `\``www` -> `%`The `Sl` -> `ns` pattern is the most concrete.The `TOP_POSITIVE_LOGITS` suggest foreign language words or technical jargon.Consider `Sl` + `ns`. Examples: `Slens` (plural of Slens?), `Lens` (if Sl is an artefact), `Client` (if `lient` is reconstructed).This is not immediately obvious.Let's focus on the most common and specific aspect: `Sl` followed by `ns`.What about `TOP_POSITIVE_LOGITS`? `ArgumentParser`, `CodeDom`, `collega`, `ersche`, `barten`, `प्रोड`, `স্ট`, `инструменты`.The foreign words strongly suggest this.Could it be detecting parts related to foreign languages or specific technical contexts where these foreign words/parts appear?Consider the combination:- `Sl` + `ns`- Foreign words (` प्रोड`, `স্ট`, `инструменты`)- Technical terms (`ArgumentParser`, `CodeDom`)The most specific pattern to explain is `Sl` followed by `ns`. The other tokens `path`, `2`, `h`, `www` don't have a strongly consistent pattern with their followers.However, the `TOP_POSITIVE_LOGITS` have *many* foreign characters and words.Perhaps the neuron is detecting specific non-ASCII characters or word roots that appear in various languages.Let's look at the tokens again:`Sl``path``2``h``www``TOP_POSITIVE_LOGITS`:`ArgumentParser` (English, technical)`ΰ` (Greek letter Upsilon, often placeholder or specific context)`collega` (Latin/Romance root)`ersche` (Germanic root)`CodeDom` (technical, .NET)`प्रोड` (Devanagari - Hindi)`স্ট` (Bengali)`инструменты` (Cyrillic - Russian for tools)`barten` (Germanic root)`⁎` (Asterisk, often symbol)This neuron seems to strongly activate on non-ASCII characters and words that are clearly not standard English or are technical.The `Sl` followed by `ns` is a specific string that might appear within such contexts.The `MAX_ACTIVATING_TOKENS` themselves might be specific prefixes or common snippets.`Sl` -- Could be a prefix`path` -- File path separator/keyword`www` -- Web prefixWhat if we combine the idea of specific prefixes/tokens with the linguistic diversity?Let's focus on `Sl` followed by `ns`. In code, this can appear in variable names or comments.The `TOP_POSITIVE_LOGITS` suggest diversity.Consider `Sl` and `ns`. What about words like "Client", "Ins" (plural), "consulting"?The neuron might be detecting segments that *could* be part of compound words or foreign terms.The presence of `http://` and `www.` in `TOP_ACTIVATING_TEXTS`, and `path` in `MAX_ACTIVATING_TOKENS`, suggests it might relate to URLs or file paths.However, the `TOP_POSITIVE_LOGITS` are overwhelmingly non-English words, Greek, Cyrillic, Devanagari, Bengali. This is a strong signal for linguistic diversity.Let's try to synthesize:- `Sl` often followed by `ns`.- Presence of foreign scripts/languages (`प्रोड`, `স্ট`, `инструменты`)

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-27b-pt-res/layer_22/width_131k
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    THEOREM
    -0.98
    Salud
    -0.97
     jardín
    -0.96
     Umfeld
    -0.93
    rotz
    -0.93
    álbum
    -0.92
     modifikasi
    -0.90
     musée
    -0.88
    pgf
    -0.88
     держать
    -0.88
    POSITIVE LOGITS
    ArgumentParser
    1.33
    ΰ
    1.23
     collega
    1.11
     ersche
    1.09
    CodeDom
    1.04
     प्रोड
    1.03
    স্ট
    1.03
     инструменты
    1.03
    barten
    1.02
    ⁎
    1.02
    Activations Density 0.038%

    No Known Activations