© 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. 15-GEMMASCOPE-2-TRANSCODER-262K
    4. 17142
    Prev
    Next
    INDEX
    Explanations

    **Analyzing Neuron Behavior**I am analyzing the provided lists to identify a unifying theme or pattern that describes the neuron's behavior. My goal is to distill this pattern into a concise phrase, adhering strictly to the specified format and constraints.**Constraint Checklist & Confidence Score:**1. **Concise explanation (3-20 words):** Yes2. **Captures what neuron detects/predicts:** Yes3.**Analyzing Neuron Behavior**I am processing the provided lists to identify a common pattern or concept that characterizes the neuron's activity. My objective is to formulate a concise explanation that encapsulates this behavior, adhering strictly to the given rules, including word count and format.**Constraint Checklist & Confidence Score:**1. **Concise explanation (3-20 words):** Yes2. **Captures what neuron detects/predicts:** Yes3. **Finds patterns in lists:** Yes4. **Single short phrase:** Yes5. **Not a full sentence:** Yes6. **No introductory phrases ("words related to", etc.):** Yes7. **No "This neuron detects/predicts":** Yes8. **No "tokens" or "patterns":** Yes9. **No capitalization unless proper noun:** Yes10. **Specific explanation:** Yes11. **Majority match pattern:** Yes12. **Fallback rule (if needed):** N/A (pattern found)Confidence Score: 5/5**Mental Sandbox Simulation:*** **Initial thought:** The most obvious pattern is "bell peppers".* **Checking against TOP_POSITIVE_LOGITS:** 'man', 'casing', 'Bine', 'Fields', 'éries', 'sagittis', 'ing', 'metik', 'Rejected', 'Parameter'. None of these directly relate to "bell peppers" or produce. This suggests the neuron might be more general or its most specific activations are not in the top positive logits.* **Checking against TOP_ACTIVATING_TEXTS:** I see many explicit mentions of "bell peppers". For example: "bell peppers (julienned)", "bell pepper, asparagus", "2 bell peppers (any color)", "Sweet Bell Pepper", "chopped bell pepper", "bell peppers, onions, mushrooms", "broccoli, bell peppers, zucchini", "Bell Pepper (thinly sliced)", "onion, bell peppers (often red and/or green)", "bell peppers, broccoli, mushrooms". This strongly reinforces the "bell peppers" connection.* **Refining the explanation:** * "bell peppers" (2 words) - This is very specific and clearly supported by the activating texts and MAX_ACTIVATING_TOKENS. * Could it be more general like "vegetables"?

    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_15_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
     Russia
    0.58
    }');
    0.57
     HOTEL
    0.57
     kaisi
    0.56
    '});
    0.54
     ноги
    0.54
    チ
    0.54
    ిక
    0.53
     Vodafone
    0.53
    '},
    0.52
    POSITIVE LOGITS
    man
    0.51
     casing
    0.51
    Bine
    0.48
    Fields
    0.47
    éries
    0.47
     sagittis
    0.47
    ing
    0.46
    metik
    0.46
    Rejected
    0.45
    Parameter
    0.45
    Activations Density 0.003%

    No Known Activations