© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    Natural Language
    Autoencoders
    NEW
    Assistant AxisNEWCircuit TracerUPDATESteerSAE EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-27B-IT
    3. 27-GEMMASCOPE-2-TRANSCODER-262K
    4. 118760
    Prev
    Next
    INDEX
    Explanations

    The core pattern across all these lists seems to be the neuron activating for the word "even" when it's used to introduce a concession or a specific framing for a refusal, particularly in the context of AI safety guidelines. The texts often say things like "even one involving..." or "even framed within...".The `MAX_ACTIVATING_TOKENS` is filled with "even".The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` has words like "one", "in", "framed", "with", "involving", "between". These often follow "even" to specify a condition or scenario that the preceding context (usually a refusal) still applies to.even followed by conditions

    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_27_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
    вайтесь
    0.37
     Polymer
    0.37
     STATEMENT
    0.37
     לג
    0.36
     HART
    0.35
     polymers
    0.35
     oni
    0.35
     \,\
    0.35
    同學們
    0.35
     justamente
    0.35
    POSITIVE LOGITS
     Leng
    0.38
    picture
    0.36
    aría
    0.36
     caractér
    0.36
    leng
    0.36
    gave
    0.36
    Serie
    0.36
     none
    0.35
    linux
    0.35
     nowhere
    0.35
    Activations Density 0.005%

    No Known Activations