© 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. Qwen3.5-4B
    3. 15-RES-MATRYOSHKA-65K
    4. 21640
    Prev
    Next
    INDEX
    Explanations

    The neuron seems to be triggered by code snippets and templating syntax, specifically looking for patterns within structural markers.Here's the breakdown:1. **MAX_ACTIVATING_TOKENS**: We see a lot of special characters and delimiters like `.`, `{{`, `${`, `+`, `%`, `"`, `,`. These are common in templating languages (like Jinja, Handlebars), shell scripting, or code snippets where parts of a string or block are marked.2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: Following these markers, we see words like `self`, `+`, `$`, `provider`, `APPLICATION`, `imag`, `s`, `response`, `addr`. These are often variable names, function calls (`self.something`), operators (`+`), or placeholders that might be filled in code.3. **TOP_ACTIVATING_TEXTS**: The examples provided strongly support the code/templating interpretation: * `self.l_name = l_name self.age = age def describe(self): ...` (Python class/method) * `"Hello, " + name ); }();` (JavaScript-like snippet) * `<tr><td>#</td><td><a href='messages/". $file .".php'>$file</a>` (PHP/HTML with variables) * `% blocktrans with provider_name=account.get_provider.name site_name=site.name %` (Templating language, like Django) * `response.status}: ${response.url}` (String interpolation or templating) * `_1_(addr) { complex _1_ addr;` (C-like structure with a variable name)4. **TOP_POSITIVE_LOGITS**: The logits are a bit more obscure, perhaps representing a latent "fuzzy" representation related to programming constructs or certain linguistic patterns that sometimes appear near code.**Conclusion:**The neuron activates when it encounters structures that resemble code, scripting, or templating language. It's learning to identify *where* dynamic content or code logic typically appears within text.**Explanation:**code delimiters 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
    oyo
    -0.07
    åij¼åij¼
    -0.05
    åı¸ç©º
    -0.05
    ä½ľç͍äºİ
    -0.05
    ugi
    -0.05
    ä¸ĭåİ»
    -0.05
    å±ķçİ°åľ¨
    -0.05
    oraf
    -0.05
    f
    -0.05
    of
    -0.05
    POSITIVE LOGITS
    sans
    0.06
    pons
    0.06
    ibern
    0.06
    rama
    0.06
    vement
    0.06
    ivor
    0.06
    ãĤĥ
    0.06
    itarian
    0.06
    anei
    0.06
    اعÙĬ
    0.06
    Activations Density 0.010%

    No Known Activations