© 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. 27-GEMMASCOPE-2-TRANSCODER-262K
    4. 120652
    Prev
    Next
    INDEX
    Explanations

    The user wants a concise explanation (3-20 words) of a neuron's behavior based on four lists. The explanation should be a phrase, not a sentence, and describe what the neuron detects or predicts by finding patterns.I need to find a common thread or pattern across the lists.1. **MAX_ACTIVATING_TOKENS**: This list is mostly empty, ending with `(-`. This is not very informative.2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: This list contains tokens like `<end_of_turn>`, numbers (`8`, `1`, `5`, `2`), single letters (`p`, `r`), short sequences (`ps`), and the word `Table`. This suggests that the neuron might be active after certain structural elements or at the end of clauses. The presence of numbers and "Table" might hint at data or enumeration.3. **TOP_POSITIVE_LOGITS**: This list contains `UserDefaults`, `мен`, `State`, `WebServer`, `Direct`, `iPad`, `Memory`, `Forget`, `Good`, `RULE`. This is a very mixed bag. `UserDefaults`, `State`, `WebServer`, `iPad`, `Memory` point towards technical/computing contexts. `Forget`, `Good`, `RULE` are general English words. `мен` is Russian.4. **TOP_ACTIVATING_TEXTS**: * "...cardiac syncope (<end_of_turn> <start_of_turn>model ```json ["diagnosis": "Cardiac Syncope",..." - Shows `<end_of_turn>` and `<start_of_turn>` markers around JSON structure.```numbers, letters, and specific words```

    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
     Cu
    0.43
     kush
    0.38
     leaned
    0.37
     hu
    0.37
    कांची
    0.37
     underwent
    0.37
     hm
    0.37
     calculer
    0.37
     alue
    0.37
     ro
    0.36
    POSITIVE LOGITS
    UserDefaults
    0.40
    мен
    0.40
    State
    0.39
    WebServer
    0.39
    Direct
    0.38
    iPad
    0.37
    Memory
    0.37
    Forget
    0.37
    Good
    0.36
    RULE
    0.36
    Activations Density 0.000%

    No Known Activations