© 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
    3. Gemma-2-27B
    5. 34-GEMMASCOPE-RES-131K
    7. 12124
    Prev
    Next
    INDEX
    Explanations

    <ctrl94>thoughtThe user wants me to explain the behavior of a neuron based on four lists: TOP_POSITIVE_LOGITS, TOP_ACTIVATING_TEXTS, MAX_ACTIVATING_TOKENS, and TOKENS_AFTER_MAX_ACTIVATING_TOKEN.The explanation needs to be concise (3-20 words) and capture what the neuron detects/predicts by finding patterns.Let's analyze the provided lists:**MAX_ACTIVATING_TOKENS**:- `double`- `export`- `of`- `ing`**TOKENS_AFTER_MAX_ACTIVATING_TOKEN**:- `1` (after `double`)- `exports` (after `export`)- `-` (after `of`)- `ing` (after `ing` - this seems self-referential or part of a longer word)**TOP_POSITIVE_LOGITS**:- `exporters`- `exporter`- `Exporter`- `export`- `exports`- `Exports`- `Export`- `Exports`- `Siempre` (Spanish for 'always' - might be an outlier or context-dependent)- `Cumulative` (Also seems a bit different)**TOP_ACTIVATING_TEXTS**:- "double exports in five years"- "export-oriented readymade garment (RMG) sector"- "Union of Banana Exporting Countries"- "export prices"- "Exporting a subset"- "Air India flights" (This seems less related to 'export' directly, but 'Air India' is in another text)- "iron ore imports prices"- "Exporting a subset"- "UAE-Bound Air India Passengers"- "activin heterodimers" (Unrelated)- "Perceptics" (Unrelated)Let's look for the most prominent pattern.The word "export" and its variations (`exports`, `exporter`, `exporters`, `Exporting`) appear very frequently in TOP_POSITIVE_LOGITS and TOP_ACTIVATING_TEXTS.- `MAX_ACTIVATING_TOKENS` has `export` and `exporter` (implied by `exporters`).- `TOKENS_AFTER_MAX_ACTIVAT ING_TOKE N` has `exports` after `export`.- `TOP_POSITIVE_LOGITS` is dominated by `export` and its forms.The text fragments strongly reinforce this:- "double exports"- "export-oriented"- "Exporting Countries"- "export prices"- "Exporting a subset"The neuron seems to be primarily focused on the concept of "exporting" or "exports".Let's consider the specific MAX_ACTIVATING_TOKENS:- `double`: Appears in "double exports".- `export`: Directly related to the main theme.- `of`: Appears in "Union of Banana Exporting Countries".- `ing`: Appears in "exporting" (from "Exporting a subset").All these bits point towards "export" and related actions/roles.The phrase should be concise and avoid meta-language.Possible explanations:- export- exports- exporting- export related words- export business- international trade- goods and services exportLooking at the list `TOP_POSITIVE_LOGITS`, the most frequent root word is `export`. Variations include `exporters`, `exporter`, `exports`.<ctrl95>export

    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_34/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
    销售
    -0.77
     imag
    -0.73
     relationships
    -0.73
     Sph
    -0.72
    omania
    -0.70
     salespeople
    -0.70
    oneofs
    -0.68
    setRight
    -0.68
    Consumer
    -0.68
    prod
    -0.68
    POSITIVE LOGITS
     exporters
    1.34
     exporter
    1.02
    Exporter
    0.96
     export
    0.93
     exports
    0.90
    Exports
    0.88
     Export
    0.83
     Exports
    0.81
     Siempre
    0.81
    Cumulative
    0.79
    Activations Density 0.007%

    No Known Activations