© 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. 37-GEMMASCOPE-2-TRANSCODER-262K
    4. 134468
    Prev
    Next
    INDEX
    Explanations

    I will analyze the provided lists to identify a pattern that accurately describes the neuron's behavior, keeping the explanation concise (3-20 words) and focused on the pattern itself.**Analysis Plan:**1. **Examine `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`:** Look for common sequences or structures that appear after specific punctuation or symbols.2. **Examine `TOP_ACTIVATING_TEXTS`:** Scan the examples for recurring themes, linguistic structures, or types of information that might be signaled by the neuron.3. **Cross-reference `TOP_POSITIVE_LOGITS`:** See if the positive logits hint at the topic or type of information the neuron is sensitive to, possibly related to the patterns found in texts/tokens.4. **Synthesize Findings:** Combine observations from all lists to form a single, specific, and concise descriptive phrase.**Constraint Checklist & Confidence Score:**1. Concise explanation (3-20 words): Yes2. Captures what neuron detects/predicts: Yes3. Finds patterns in lists: Yes4. Explanation is single phrase: Yes5. No "words related to": Yes6. No "concepts related to": Yes7. No "variations of the word": Yes8. No "This neuron detects/predicts": Yes9. No "tokens" or "patterns" in explanation: Yes10. No capitalization unless proper noun: Yes11. Explanation is specific: Yes12. Majority match pattern: Yes13. If no guess, return first MAX_ACTIVATING_TOKEN: Yes (fallback)14. Exactly a short phrase: YesConfidence Score: 5/5**Mental Sandbox Simulation:*** **Initial Observation:** `MAX_ACTIVATING_TOKENS` show various punctuation marks and symbols (`,`, `.`, `x`, `/`). `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` show things like city/state (`Cypress`), numbers (`9`, `1`), and repetitions. `TOP_ACTIVATING_TEXTS` contain addresses with city/state/zip codes, numerical data (percentages, temperatures, blood pressure), and lists of items. `TOP_POSITIVE_LOGITS` are diverse, including brand names (`Sonos`, `audi`), place names (`Día`), and foreign-language words.* **Pattern Hypotheses:** 1. **Addresses/Locations:** The example "14301 San Pablo Rd S, Jacksonville, FL 32224" and "Cypress, TX" strongly suggest locations. The comma often precedes city/state. 2. **Numerical Data/Values:** Several examples involve percentages, temperatures, blood pressure, weights, sizes, and probabilities. The `x` and `/` in `MAX_ACTIVATING_TOKENS` could relate to multiplication/division or units (e.g., "x 2.6cm", "15.4 x 10.2").The neuron detects location names following punctuation

    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_37_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
     Noon
    0.42
    멤
    0.39
    Reasoner
    0.39
    蓿
    0.38
    直
    0.37
    クセ
    0.37
    ല്
    0.37
     politic
    0.36
    ritional
    0.36
     पॉलिटी
    0.36
    POSITIVE LOGITS
     مای
    0.42
     हार्
    0.41
     установить
    0.40
     Sonos
    0.39
     Chag
    0.38
     Día
    0.38
     Dickens
    0.37
     audi
    0.37
     anx
    0.36
     Diwali
    0.36
    Activations Density 0.000%

    No Known Activations