© 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. 46-GEMMASCOPE-2-TRANSCODER-262K
    4. 141976
    Prev
    Next
    INDEX
    Explanations

    Limited liability companiesFinal Answer structure: Method X is used. Reason. Explanation.Let's redo the methods in order with the right mindset.**Neuron 5****Method 1: MAX_ACTIVATING_TOKENS**`of`, `a`, `Limited`, `:**`, `limited`, `a`, `a`, `-`, `Limited`, `limited`These tokens are a mix of common words, punctuation, and capitalization variations of "limited". They do not represent a single, specific token or a clear variation of one. Method 1 fails.**Method 2: TOKENS_AFTER_MAX_ACTIVATING_TOKEN**`limited`, `*`, `Liability`, `Limited`, `liability`, `company`, `company`, `Limited`, `Company`, `liability`Let's check for patterns:- Do they all start with the same letter? No (l, *, c).- Is there another specific pattern like "word ending in -ing"? No.- These tokens strongly suggest the concept of "limited liability company". However, the instruction is to find a pattern *in the tokens themselves*, like "all start with 'w'" or "'s word]". The concept is present, but not a simple text pattern *across all tokens* in the list as defined by the examples (e.g., 's word, ending -ing, number 8). Method 2 fails to find a simple textual pattern across all tokens for its specific conditions.**Method 3: TOP_POSITIVE_LOGITS**`SPD`, `ość`, `ood`, `juč`, `ISD`, `unistd`, `Adopt` (the prompt shows these, but excludes foreign characters for the most part for this specific case - the actual output was `SPD`, `XYZ`, `飛`, `ość`, `ood`, `juč`, `ISD`, `unistd`, `Adopt`)Looking at the provided TOP_POSITIVE_LOGITS: "apple", "orange", "pineapple" etc. for Neuron 2. That was clear.For Neuron 5, the list is: `SPD`, `XYZ`, `飛`, `ość`, `ood`, `juč`, `ISD`, `unistd`, `Adopt`.These are highly varied: acronyms, foreign characters, technical terms. There isn't a clear, concise description (1-3 words) that applies to *all* of them. "Non-standard tokens" might be too vague. "Foreign characters" only applies to some. "Mix of characters" is also vague. The instructions emphasize specificity. Method 3 fails to find a specific, brief pattern description.**Method 4: TOP_ACTIVATING_TEXTS**The texts are about:- "Ltd."- "Designated Activity Company (DAC)"- "limited company"- "Sociedad Limitada (SL) - Limited Liability Company"- "Limited Company (Ltd)"- "GmbH (Gesellschaft mit beschränkter Haftung - limited liability company)"- "Limited legal protection"- "Sociedad Limitada - Limited Company"

    np_max-act-logits · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-it/transcoder_all/layer_46_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
     prodotti
    0.47
     produkter
    0.46
     продукти
    0.46
     BEACH
    0.45
     lawful
    0.45
     Produkte
    0.44
    产品的
    0.44
     beach
    0.43
     कॉर्
    0.43
     productos
    0.43
    POSITIVE LOGITS
    SPD
    0.41
     उड़
    0.38
    XYZ
    0.38
    飛
    0.37
    ość
    0.36
    ood
    0.35
    juč
    0.35
     ISD
    0.35
    unistd
    0.35
     Adopt
    0.35
    Activations Density 0.003%

    No Known Activations