© 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. 120739
    Prev
    Next
    INDEX
    Explanations

    publication title followed by a colon.This is exactly what is seen in `TOP_ACTIVATING_TEXTS`:- "Science Focus:"- "HistoryExtra:"- "Science Focus Magazine:" (here Magazine follows, not :)- "BBC Science Focus Magazine:"- "BBC Science Focus Magazine:"- "ScienceFocus:"- "Tennis World USA:"- "Interesting Engineering:"- "Nutraceutical Business Review:"- "The Poland Guide:"The pattern of `MAX_ACTIVATING_TOKENS` like `Focus`, `World`, `Engineering`, `Review`, `Guide` followed by `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` like `:**` or `Magazine` strongly supports this.Let's refine the explanation based on this. The core pattern is recognizing these publication names, often followed by punctuation.Try to be more specific than just "publication titles."The `MAX_ACTIVATING_TOKENS` are often *parts* of publication names.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are separators or type extensions.It's about names that look like magazine/news/journal titles.Phrases:- Publication titles with colons- Magazine and journal titles- Titles of publicationsLet's consider *how specific* it needs to be.The phrase needs to capture the "what" the neuron detects.The most consistent *visual/textual* pattern is the format of these publication titles.Re-evaluating `MAX_ACTIVATING_TOKENS`:Focus, Extra, Focus, Focus, Focus, Focus, World, Engineering, Review, Guide.These are *words that commonly appear at the beginning or as key parts of publication titles*.Re-evaluating `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`::**, Magazine, :, :, :, USA, :, :, :.These are *common suffixes or delimiters following these publication-like words*.The `TOP_ACTIVATING_TEXTS` confirm this, showing strings like "Science Focus:", "Tennis World USA:", "Interesting Engineering:".So the neuron detects the *structure* or *form* of these publication titles, especially when followed by a colon.Possible explanations:- Publication title format- Magazine title format- Names followed by a colon- Recognize publication titles- Publications followed by separators- Title followed by colon"Publication titles followed by separators" feels accurate and within the word count."magazine title followed by colon" focuses too much on "magazine" and "colon" specifically, when "USA" and "Magazine" are also seen.Let's look at the prompt's examples for inspiration:- "predicts conjunctions"- "detects dates"- "identifies dates"- "detects years"- "identifies dates"These are direct about the *type of information*.**publication titles followed by separators**

    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
    古典
    0.42
     berühm
    0.41
     célèbre
    0.40
     famously
    0.40
    हूर
    0.40
     বিখ্যাত
    0.39
     jauh
    0.39
     bezpečnost
    0.39
     جميل
    0.38
     証明
    0.38
    POSITIVE LOGITS
    Diagnostic
    0.36
    াব
    0.36
    Introduce
    0.35
    噵
    0.34
    Runtime
    0.32
    Projection
    0.32
    tried
    0.32
    UNDE
    0.32
    َات
    0.32
     Projection
    0.32
    Activations Density 0.002%

    No Known Activations