© 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-2-27B
    3. 22-GEMMASCOPE-RES-131K
    4. 11165
    Prev
    Next
    INDEX
    Explanations

    explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"}, {"type": "phrase", "text": "explains"} In this case, the neuron seems to be associated with words that *educate* or *explain*. The `MAX_ACTIVATING_TOKENS` `educ` combined with `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` ``ates` strongly points to `educates`. The `TOP_ACTIVATING_TEXTS` also contain examples like "booklet that educates". Although many other texts are about technical problems, the core signal seems to be about providing information or explanation.Therefore, the explanation is "explains".Final answer needs to be 3-20 words. "explains" is one word. Let's re-check if I can capture more.The texts are about *how-to*, *problem-solving*, *information lookup*. This implies explanation or teaching."explains how to" would be good, but "how to" isn't directly in the tokens.Given the strong signal of `educates` and the nature of the other texts (problem descriptions, guides), "explains" is the most direct and specific pattern. explains

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-27b-pt-res/layer_22/width_131k
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     should
    -2.48
     in
    -2.36
     There
    -2.31
     on
    -2.19
     When
    -2.16
     at
    -2.09
     After
    -2.02
    が
    -1.99
    on
    -1.95
     get
    -1.95
    POSITIVE LOGITS
    </em>
    2.81
    </i>
    2.20
    笮
    2.11
    </b>
    2.08
    ;</
    2.03
     ausein
    2.03
    merkmale
    2.02
     dezelve
    2.02
    ,</
    1.99
    "));
    
    1.95
    Activations Density 0.003%

    No Known Activations