© 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 EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-27B-IT
    3. 16-GEMMASCOPE-2-TRANSCODER-262K
    4. 216848
    Prev
    Next
    INDEX
    Explanations

    analyzing punctuation, special characters, and followed words.````TOP_ACTIVATING_TEXTS`:- Contains phrases like "Extremely Toxic", "Therapies (ACTs)", "architectural optimizations", "wuxia game", "dimensional anomalies", "key mechanic. Holding metal objects during a storm dramatically increases your chance of being struck. * **Height Matters ...", "predicate 'nrel_concept_entity'", "I, as an AI, am still "processing"...".**Analysis:**1. **MAX_ACTIVATING_TOKENS**: Includes punctuation (`!`, `.`, `(`), technical terms (`optimizations`), foreign characters (`寺`), common words (`or`, `Height`), and an exclamation mark.2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: Shows words like `This`, `(`, `.`, `optimizations`, `power`, `Matters`, `second`, `It`. This suggests sequential relevance or context following the activated token. The appearance of `Matters` after `or` and `second` seems a bit scattered, but `power` after `寺` might be relevant in a context. `This(` after `!` or `optimizations(` after `optimizations` hints at structural or contextual continuation.3. **TOP_POSITIVE_LOGITS**: Primarily non-English words or specific technical/code-like tokens (`済み`, `の後`, `廃`, `]()`, `עי`, `лишком`, `的问题`, `ності`, `পেতে`, `買って`). This heavily suggests a neuron that reacts to or activates on multi-lingual content, specific foreign scripts, or specific technical/code syntax.4. **TOP_ACTIVATING_TEXTS**: * "Extremely Toxic" - suggests classification or judgment. * "Therapies (ACTs)" - includes acronyms and parenthetical content. * "architectural optimizations" - technical jargon. * "wuxia game" - specific cultural/genre context. * "dimensional anomalies. ... * **Height Matters..." - mentions concepts related to physics, height, and danger. * "predicate 'nrel_concept_entity'" - highly technical, formal logic/data structure syntax. * AI processing - meta-commentary on AI capabilities.**Pattern Identification:*** The `TOP_POSITIVE_LOGITS` list is dominated by non-English words and specific technical/code elements.* `MAX_ACTIVATING_TOKENS` shows a mix, but `寺` (temple) might hint at specific contexts. Punctuation is also common.* The `TOP_ACTIVATING_TEXTS` include a wide range of topics and styles, but several involve technical descriptions (optimizations, predicates), specific genre contexts (wuxia), or academic/scientific discussions (anomalies, height, therapies). Some texts also contain specific formatting like parentheses, asterisks, and code-like structures.The strongest signal comes from `TOP_POSITIVE_LOGITS` and the technical/multi-lingual aspects seen in `TOP_ACTIVATING_TEXTS`.multi-lingual technical structures and foreign characters

    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_16_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
    e
    0.64
    е
    0.48
     };
    0.48
    ulations
    0.47
    ර්
    0.47
    О
    0.47
    E
    0.46
    c
    0.46
    };
    0.45
    Ма
    0.44
    POSITIVE LOGITS
    済み
    0.46
    の後
    0.46
    廃
    0.45
    ]()
    0.45
    עי
    0.45
    лишком
    0.45
    の問題
    0.44
    ності
    0.44
     পেতে
    0.44
    買って
    0.44
    Activations Density 0.000%

    No Known Activations