© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it
    Neuronpedia
    Jacobian LensNEW
    Natural Language
    Autoencoders
    MAY 26
    Assistant AxisJAN 26Circuit TracerAPR 26SteerSAE EvalsExportsinterp-engineNEW Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-27B-IT
    3. 16-GEMMASCOPE-2-TRANSCODER-262K
    4. 219085
    Prev
    Next
    INDEX
    Explanations

    The provided lists show evidence that the neuron is activated by common words and parts of words, often followed by words that describe actions, states, or continuations.1. **MAX_ACTIVATING_TOKENS**: `will`, `final`, `and`, `captures`, `(`, `ulate`, `were`, `were`, `,`, `to`2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `congregate`, `das`, `difficult`, `the`, `see`, `effective`, `abused`, `abused`, `we`, `understand`3. **TOP_POSITIVE_LOGITS**: `體的`, `饮`, `న్స్‌`, `খাওয়`, `嘚`, `getBody`, `屶`, `Body`, `รั่ง`, `➎`4. **TOP_ACTIVATING_TEXTS**: Contains examples like "Corvids **will congregate**", "o 's' e o 'z' no **final das** sílabas", "slow to train and **difficult to** parallelize", "If it's a **capture, use** `x`", "**Analyze situations:** **See** opportunities", "You **were abused**", "your father for a long time and had multiple pregnancies and abortions. When you **were** pregnant".Let's look for patterns:* **`will congregate`**: Appears in MAX_ACTIVATING_TOKENS (`will`) and TOKENS_AFTER_MAX_ACTIVATING_TOKEN (`congregate`). This is a clear phrase in the activating text.* **`final das`**: `final` is in MAX_ACTIVATING_TOKENS, `das` is in TOKENS_AFTER_MAX_ACTIVATING_TOKEN. Part of "final das sílabas".* **`and difficult`**: `and` is in MAX_ACTIVATING_TOKENS, `difficult` is in TOKENS_AFTER_MAX_ACTIVATING_TOKEN. Part of "slow to train and difficult to parallelize".* **`captures the`**: `captures` is in MAX_ACTIVATING_TOKENS,

    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
    Negative Logits
    я
    0.51
    il
    0.44
     이런
    0.43
     религи
    0.43
    ie
    0.42
     गेल
    0.42
    aid
    0.41
    хий
    0.41
    y
    0.41
    शिक
    0.40
    POSITIVE LOGITS
    體的
    0.51
    饮
    0.49
    న్స్‌
    0.48
     খাওয়
    0.48
    嘚
    0.47
    getBody
    0.46
    屶
    0.46
     Body
    0.45
    รั่ง
    0.45
    ➎
    0.45
    Activations Density 0.000%

    No Known Activations