© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it
    Neuronpedia
    Jacobian LensNEW
    Natural Language
    Autoencoders
    MAY 26
    Assistant AxisJAN 26Circuit TracerAPR 26SteerSAE EvalsExportsinterp-engineNEW Community BlogPrivacy & TermsContact
    1. Home
    2. Qwen3.5-4B
    3. 15-RES-MATRYOSHKA-65K
    4. 22371
    Prev
    Next
    INDEX
    Explanations

    Let's break down the neuron's behavior based on the provided lists.1. **MAX_ACTIVATING_TOKENS**: The most prominent token repeated here is "ion". However, "accession" stands out as a specific word. We also see "IP" and "CKET".2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: * After "accession", we see "Ranges", "numbers", "number". * After "contains", we see "Pro". * After "IP", we see "in" and "of". * After "CKET", we see "NO".3. **TOP_POSITIVE_LOGITS**: This list contains words like "roid", "ampire", "plo", "protein", "loid", "leton", "star". These seem somewhat diverse but might relate to scientific/technical terms or possibly classifications.4. **TOP_ACTIVATING_TEXTS**: * "accessionRanges" followed by "D03833-D03973; D40780" (numbers/ranges) * "accession numbers: 6KR2 and 6KR3" (scientific identifiers) * "accession numbers KT279554 to KT279561" (scientific identifiers) * "accession number KT072629" (scientific identifier) * "GeneAccession \#" followed by codes (scientific/gene identifiers) * "IP address report" followed by an IP address. * "DOCKET NO. A-2395-17T1" (document identifier)**Pattern Identification**:The strongest recurring theme is around "accession" followed by identifiers like numbers, ranges, codes, or specific IDs. This strongly suggests referencing some kind of catalog, database, or filing system. The "IP address" and "DOCKET NO." also point towards identification or reference numbers. The "ion" suffix is common to many technical terms, fitting in scientific contexts.**Formulating the Explanation**:The neuron is clearly associating the word "accession" (and similar concepts like "IP" or "DOCKET") with identifying numbers, codes, or ranges.Considering the rules:* Concise (3-20 words)* Find patterns* No "tokens", "patterns", "detects", "predicts"* Specific* Capitalize only proper nouns (none here)The most specific and recurring pattern is around accession numbers and related identifiers.Let's try some phrases:* accession numbers and identifiers* references to accession numbers* accession numbers or identifiers* catalog accession numbers* database accession identifiers"accession numbers or identifiers" seems to cover "accession numbers", "accession ranges", "accession number", "GeneAccession #", "IP address", and "DOCKET NO.", all of which are specific types of identifiers. It's also 4 words, within the limit.Final check: is it specific? Yes, it points to accession numbers and other identifiers. Does it avoid forbidden words? Yes. Is it concise? Yes.Final Answer:accession numbers or identifiersaccession numbers or identifiers

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    decoderesearch/qwen-3.5-saes/qwen-3.5-4b
    Prompts (Dashboard)
    16,384 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Negative Logits
    оÑĩи
    -0.07
     reckon
    -0.07
    fore
    -0.06
    ÙĪØ§Ø¡
    -0.06
    ynos
    -0.06
    reas
    -0.06
    åľ¨å¤ĸ
    -0.06
     Statt
    -0.05
    端起
    -0.05
     perspek
    -0.05
    POSITIVE LOGITS
    roid
    0.06
    ampire
    0.06
    plo
    0.05
    å¾ħãģ¡
    0.05
    rotein
    0.05
    å¾ħ
    0.05
    çĨŁ
    0.05
    loid
    0.05
    leton
    0.05
    tar
    0.05
    Activations Density 0.001%

    No Known Activations