© 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-3-12B
    3. 24-GEMMASCOPE-2-RES-16K
    4. 10093
    Prev
    Next
    INDEX
    Explanations

    The neuron appears to activate around the concept of **qualifying leads** or the **process of becoming a qualified lead**, with connections to business and sales terminology like "prospect" and "companies". It also seems to touch upon the idea of flagging or identifying something, as suggested by "mark". The presence of terms like "disinformation" might indicate it's also identifying topics that require careful qualification or distinction.However, the `TOP_POSITIVE_LOGITS` list contains non-Latin characters (س, க்கு, ب, ০, ل) and some Italian/other character sequences (iono, に至, ionario). This is a strong indication that the neuron might be associated with languages other than English, or is perhaps picking up on specific linguistic features or transliterations related to certain languages or technical terms.Given the prompt's request to find *patterns* and avoid simply listing tokens, and specifically focusing on what the neuron "detects or predicts by finding patterns in lists":Let's try to find a unifying theme, considering the strong signals from `MAX_ACTIVATING_TOKENS` and the less clear, but present, `TOP_POSITIVE_LOGITS`.`prospect` and `companies` in `MAX_ACTIVATING_TOKENS` point to business contexts.`mark` could relate to 'marking' something for attention or evaluation.`disinformation` implies identifying problematic content.The combination of `prospect` -> `qualified` lead in the `TOP_ACTIVATING_TEXTS` is a very strong signal.The `TOP_POSITIVE_LOGITS` (س, க்கு, ب, ০, ل, iono, に至, ionario) are highly specific and non-English. This suggests the neuron might be sensitive to specific linguistic roots, phonemes, or character sequences that appear in certain languages or technical terms. For example:- `iono` and `ionario` could relate to words ending in "-ion" or words borrowed from Romance languages.-

    np_acts-logits-general · gemini-2.5-flash-lite

    This neuron activates on bibliographic‐citation elements (e.g. author names, numbers, journal volumes/pages) in reference lists.

    oai_token-act-pair · o4-miniTriggered by @jyhe0408
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-12b-pt/resid_post/layer_24_width_16k_l0_medium
    Prompts (Dashboard)
    392,802 prompts, 256 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     parcels
    0.86
    тым
    0.82
    鬚
    0.79
     moneys
    0.79
    졔
    0.77
    ᅣ
    0.75
     filets
    0.74
     தின
    0.73
     cladding
    0.73
     fringes
    0.71
    POSITIVE LOGITS
    س
    0.82
     
    0.75
    res
    0.72
    ০
    0.72
    elio
    0.71
    க்கு
    0.70
    ますが
    0.70
    ل
    0.70
    us
    0.68
    ب
    0.67
    Activations Density 0.001%

    No Known Activations