© 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-3-27B-IT
    3. 16-GEMMASCOPE-2-TRANSCODER-262K
    4. 214664
    Prev
    Next
    INDEX
    Explanations

    Ok, here are the lists:MAX_ACTIVATING_TOKENS:plungewriteJustherdeemed0OkSuggestionsSuggestionsarseTOKENS_AFTER_MAX_ACTIVATING_TOKEN:cutinsaybestessentialcctaforunaTOP_POSITIVE_LOGITS:Người值的v敎Basketball'>"+Honest死JudgesতেTOP_ACTIVATING_TEXTS:- . Once the disc bites, gradually bring the grinder to a 90-degree angle. *Never* plunge cut into concrete or tile - it will likely shatter the disc.Ok, I understand. I need to analyze the provided lists to identify a pattern or concept that describes the neuron's behavior within the 3 to 20-word limit, avoiding specific phrases like "this neuron detects" and focusing on the detected characteristic itself.Let's break down the provided lists:* **MAX_ACTIVATING_TOKENS**: `plunge`, `write`, `Just`, `her`, `deemed`, `0`, `Ok`, `Suggestions`, `Suggestions`, `arse`* **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `cut`, `in`, `say`, `best`, `essential`, `cc`, `ta`, `for`, `una`* **TOP_POSITIVE_LOGITS**: `Người`, `值的`, `v`, `敎`, `Basketball`, `'>"+`, `Honest`, `死`, `Judges`, `তে`* **TOP_ACTIVATING_TEXTS**: This list contains example texts where the neuron exhibits strong activation. I need to look for common themes or structures in these texts, especially around the `MAX_ACTIVATING_TOKENS`.Let's examine `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` together first:* `plunge` -> `cut` (e.g., "plunge cut")* `write` -> `in` (e.g., "write in Chinese")* `Just` -> `say` (e.g., "Just say")* `her` -> `best` (potentially part of a phrase like "her best")* `deemed` -> `essential` (e.g., "deemed essential")* `Ok` -> `for` (e.g., "Ok for")These pairs (`plunge` + `cut`, `deemed` + `essential`) suggest the neuron might be sensitive to specific collocations or phrases where a particular word is immediately followed by a related word.Now let's look at the `TOP_ACTIVATING_TEXTS` for confirmation and broader context:1. "...*Never* **plunge** cut into concrete or tile..." - Confirms `plunge cut`.2. "...Please **write** in Chinese language." - Confirms `write in`.3. "...**Just** say 'Use /start to begin!'..." - Confirms `Just say`.4. "...skills **deemed** essential for ruling." - Confirms `deemed essential`.5. "...**Suggestions** for hatyai travel..." - `Suggestions` is followed by `for`. This fits the pattern of a specific word followed by another word, though `for` wasn't in the `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list for `Suggestions` in this particular input (it was `for` after `Ok`). However, `Suggestions` appears twice, indicating its importance.6. "...**Ok**, here are some suggestions..." - `Ok` followed by `here` (not in the list after). This might be an exception or less strongly correlated.7. The text referring to "800cc era" has `0` in the `MAX_ACTIVATING_TOKENS` list. It's hard to link this directly to a subsequent token without the full context, but it might relate to numerical values.The `TOP_POSITIVE_LOGITS` list (`Người`, `值的`, `v`, `敎`, `Basketball`, `'>"+`, `Honest`, `死`, `Judges`, `তে`) seems quite diverse and not directly tied to the specific token pairings observed. It suggests the neuron's *activation* is high for these, but the *reason* seems to stem more from the structural patterns in `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`.Identifying specific word pairings or sequences

    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_16_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
    َل
    0.53
     acumen
    0.52
     evid
    0.52
     oxidation
    0.51
     оки
    0.51
     immunization
    0.50
     infallible
    0.50
     intuit
    0.49
     enchantment
    0.48
     amaze
    0.47
    POSITIVE LOGITS
     Người
    0.48
    值的
    0.46
    v
    0.46
    敎
    0.46
     Basketball
    0.44
    '>"+
    0.44
     Honest
    0.44
    死
    0.44
     Judges
    0.43
    তে
    0.43
    Activations Density 0.000%

    No Known Activations