© 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. 37-GEMMASCOPE-2-TRANSCODER-262K
    4. 126605
    Prev
    Next
    INDEX
    Explanations

    "anime" or "Japanese culture" related content or prompts.Let's break down how I arrived at this explanation:* **TOP_POSITIVE_LOGITS**: This list is a primary indicator. It's filled with terms like "Hokkaido", "kimono", "Japanese", "일본" (Korean for Japan), "জাপান" (Bengali for Japan), "Tokyo", "Osaka", and "जापान" (Hindi for Japan). This strongly points to a focus on Japanese locations, culture, and language.* **TOP_ACTIVATING_TEXTS**: This list provides context. I see phrases like: * "...an anime story about a girl in a new world..." * "...writing style a professionally edited bestselling author and popular anime..." * "...as a anime<end_of_turn>..." * "...a slutty celestia ludenberg from danganronpa..." (Danganronpa is a Japanese visual novel/anime franchise). * "...You're an anime girls, romance me..." These texts explicitly mention "anime", "anime story", "anime character", and characters/terms strongly associated with anime or Japanese pop culture ("celestia ludenberg", "danganronpa").* **MAX_ACTIVATING_TOKENS** and **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: While these lists seem more fragmented and contain special characters or grammatical markers (`'`, `the`, `<start_of_turn>`, `:**`, `d`), their context within the activating texts is important. For instance, `<start_of_turn>` often signals the beginning of a user prompt or a conversational turn, which, in conjunction with the other lists, frequently leads to requests related to anime.**Synthesis**:The strong presence of Japanese cultural terms in `TOP_POSITIVE_LOGITS` combined with the explicit mentions of "anime" and anime-related franchises in `TOP_ACTIVATING_TEXTS` leads to the conclusion that this neuron is associated with detecting or predicting requests or content related to anime and Japanese culture. The other lists, while less direct, support this by showing how these concepts appear in conversational turns or specific text continuations.Therefore, the most concise and accurate explanation is: **anime or Japanese culture requests/content**.anime or Japanese culture requests/content

    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_37_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
     Tamara
    0.40
    பாத
    0.38
    膦
    0.38
    Diss
    0.37
     Pur
    0.37
     Piet
    0.37
     Pagan
    0.37
    ഴ്
    0.37
    Spr
    0.36
     Alej
    0.36
    POSITIVE LOGITS
     Hokkaido
    0.71
     kimono
    0.68
     Japanese
    0.66
     일본
    0.65
     জাপান
    0.64
     Shin
    0.63
     Tokyo
    0.62
     Osaka
    0.61
     kimchi
    0.61
     जापान
    0.61
    Activations Density 0.000%

    No Known Activations