© 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. 10-GEMMASCOPE-2-TRANSCODER-262K
    4. 46423
    Prev
    Next
    INDEX
    Explanations

    **Analyzing Neuron Behavior**I have analyzed the provided lists to identify the pattern associated with this neuron.* **MAX_ACTIVATING_TOKENS**: Consistently shows "designs".* **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: Shows tokens like ".", "that", "inspired", "for", "or", "and".* **TOP_POSITIVE_LOGITS**: Shows a mix of characters/letters from different scripts, which might point to a broad, potentially cross-lingual aspect or simply noise.* **TOP_ACTIVATING_TEXTS**: Contains phrases like "future journeys, and designs", "Pokémon with designs that evoke", "designs inspired by Japanese animation/comics", "sketching designs for a particularly elegant building", "Look for fun designs or colors", "intricate designs", "create specific designs and shapes", "outlandish designs", "artist behind the *Alien* designs", "generates designs".The overwhelming presence of "designs" in MAX_ACTIVATING_TOKENS, combined with the context from TOP_ACTIVATING_TEXTS, clearly indicates that this neuron is strongly associated with the concept of "designs". The tokens following "designs" suggest the context of what these designs are *for* or *inspired by*.The TOP_POSITIVE_LOGITS are diverse and don't form clear English words, suggesting the neuron might not be tied to specific English words in positive logits, but rather to a conceptual understanding, or perhaps it's capturing very abstract, cross-lingual features. Given the rules to be specific and avoid abstract terms if possible, and to focus on patterns, the most dominant and specific signal is "designs".Let's check the rules:- Concise (3-20 words): Yes.designs

    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_10_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
     nyingi
    0.59
     binge
    0.59
     HIIT
    0.58
     memper
    0.57
     HIPAA
    0.53
     gripper
    0.52
     ноутбу
    0.52
    MSFont
    0.51
     получать
    0.51
     ciné
    0.51
    POSITIVE LOGITS
    O
    0.73
    Sh
    0.71
    ج
    0.71
    1
    0.70
    На
    0.69
    За
    0.67
    Z
    0.66
    گ
    0.66
    We
    0.64
    ست
    0.64
    Activations Density 0.012%

    No Known Activations