© 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. 16858
    Prev
    Next
    INDEX
    Explanations

    The pattern seems to be related to code elements and parameters.Let's look for common threads:- `sum` is followed by `([` (likely a list literal).- `Sink` is followed by `().` (likely method/constructor call).- `-d` could be a flag/parameter, and `iscrim` is after it.- `_field` is followed by `</` which is common in HTML/XML tags.- `ViewModel` is followed by `Summary`.- `-values` is followed by `{:`.- `1` is followed by `0`.- `p` is followed by `(file`.- `(Item` is followed by `Of`.- `x` is followed by `array`.The `TOP_POSITIVE_LOGITS` are mostly non-English characters or specific code-related terms.The `MAX_ACTIVATING_TOKENS` also contain code-related identifiers and symbols like `_field`, `ViewModel`, `(Item`, `sum`, `Sink`, `-d`, `-values`.The pattern appears to be a mix of code tokens, function/method calls, and potentially list/iterable items. There's a mix of identifiers, special characters, and fragments of code structure. The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` often represent the start of a code construct (like `([`, `().`, `{:`, `(file`) or closing tags (`</`).Given the mix of code identifiers, programming symbols, and the following tokens often initiating code blocks or syntax, a good explanation would be related to code structure or programming elements.Let's refine based on "finding patterns in lists" and being "specific".code elements and structure

    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
    Not in Any Lists

    No Comments

    Negative Logits
    ereg
    -0.06
    ä¸Ģå¤ķ
    -0.05
    olda
    -0.05
    ëĭ¹
    -0.05
    åĪ©çĽĬ
    -0.05
    anger
    -0.05
    饮
    -0.05
    perc
    -0.05
    eliac
    -0.05
    kest
    -0.05
    POSITIVE LOGITS
    ç«£
    0.06
    人ä¸Ń
    0.06
    äºĭä»¶ä¸Ń
    0.06
    reo
    0.05
     poj
    0.05
    -es
    0.05
    Agenzia
    0.05
     Pequ
    0.05
    иж
    0.05
    ÙģÙĪ
    0.05
    Activations Density 0.013%

    No Known Activations