© 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. 2-GEMMASCOPE-2-TRANSCODER-262K
    4. 108079
    Prev
    Next
    INDEX
    Explanations

    The neuron seems to be responding to tokens that begin or contain the 'z' sound, particularly when followed by specific characters or word structures. The presence of Cyrillic characters (like 'зе') and the common follow-up element 'я' (which often follows 'z' in Slavic languages) suggests a focus on these phonetic or orthographic patterns. The repetition of `ZIE` followed by `LE` might indicate a specific English or technical context, but the dominant pattern involves the 'z' sound.The most striking and consistent pattern is the presence of tokens starting with 'z' or 'zi', often associated with Slavic languages given the `зе` and `я` context.Let's re-evaluate based solely on the letters in MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN for a simpler pattern.MAX_ACTIVATING_TOKENS:зеzizziZIEZIEZIEɕTOKENS_AFTER_MAX_ACTIVATING_TOKEN:я .** ienie zaz 2 LE LE LE en 8Patterns to consider:1. **Starts with 'z'**: `зе`, `zi`, `z`, `zaz`. This is strong.2. **Specific letter combinations**: `zi`, `ze`. `ZIE` is also present.3. **Followed by specific tokens**: `зе` -> `я`, `zi` -> `zaz`. `ZIE` -> `LE`.The common thread is indeed the 'z' sound or letter

    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_2_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
    ри
    1.66
    ן
    1.41
    ל
    1.38
    るので
    1.35
    т
    1.32
    ت
    1.32
    י
    1.29
    ຜະລິດຕະພັນ
    1.20
    ш
    1.20
    ﻧ
    1.20
    POSITIVE LOGITS
     a
    1.93
    ind
    1.20
     at
    1.17
    ach
    1.17
    ade
    1.17
     (
    1.16
     not
    1.14
    ans
    1.12
    aya
    1.09
     about
    1.07
    Activations Density 0.000%

    No Known Activations