© 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. Olmo-3-1025-7B
    3. 16-RES-MATRYOSHKA-65K
    4. 1291
    Prev
    Next
    INDEX
    Explanations

    Let's break down the provided lists to pinpoint the neuron's behavior.1. **MAX_ACTIVATING_TOKENS**: * `class` * `https` * `=http` * `https` * `https` * `https` * `class` * `http` * `Yii` * `http` This list shows tokens that strongly trigger the neuron. We see a mix: `https` and `http` (common in URLs), `class` (common in HTML/code), and `Yii` (a PHP web framework).2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: * `="` (after `class`) * `://` (after `https`) * `://` (after `https`) * `://` (after `https`) * `://` (after `https`) * `://` (after `https`) * `="` (after `class`) * `://` (after `http`) * `::` (after `Yii`) * `://` (after `http`) This shows what follows these tokens. * `class` is followed by `="`. This is a strong indicator of HTML attributes. * `http`/`https` are followed by `://`, which is standard for URLs. * `Yii` is followed by `::`, common in PHP for static calls or namespaces, often related to framework features.3. **TOP_POSITIVE_LOGITS**: * `Int` * `coded` * `/devices` * `spiel` * `ministries` * `stell` * `enso` * `)+` * `pawn` * `mighty` These are the most likely *words* associated with the neuron's output, not necessarily directly from the input text. They seem quite diverse and don't immediately form a clear theme *unless* they represent classifications or labels. However, the pattern from the other lists is stronger.4. **TOP_ACTIVATING_TEXTS**: * `class="group-title">` * `https://localhost:44300/account/login` * `https://localhost:44300/Account/Login?returnUrl=` * `<a href=http://cmsdoc.cern.ch/swdev/viewcvs/viewcvs.cgi/CMSSW/Reco` * `// https://tools.ietf.org/html/rfc7292#appendix-B.1` * `following links: <https://trace.ncbi.nlm.nih.gov/Traces/sra/?study=SRP120033>` * `v3.pdf](https://pjreddie.com/media/files/papers/YOLOv3.pdf)` * `GitHub: [https://github.com/pjreddie/darknet]` * `<th class="wsite-com-checkout-list-item"` * `http://www.mysite.com/myfolder/page1.php` * `http://www.mysite.com/myfolder/comman.php?cat=page1` * `Yii::app()->getController()->id` * `http://www.astroalert.su/files/m32_2011-10-01_pnv_elenin_summ3.png` This confirms the presence of: * HTML-like structures (`class="..."`, `href=http...`) * URLs (`http://...`, `https://...`) * PHP/Yii specific code constructs (`Yii::app()->...`)**Pattern Identification**:The most consistent and specific patterns are:1. `class` followed by `="` (HTML attributes).2. `http` or `https` followed by `://` (URLs).3. `Yii` followed by `::` (PHP framework/namespace).These all point towards web-related code, specifically web addresses and code structures commonly found in web development. The `TOP_POSITIVE_LOGITS` seem less directly related to the input's structural patterns and might be more about conceptual associations or class labels derived from the overall context of such inputs.The core of the neuron's detection seems to be related to these web components.**Concise Explanation**:This neuron detects components related to web addresses and code structures common in web development.* `http`, `https` followed by `://`* `class` followed by `="`* `Yii` followed by `::`Let's try to synthesize this into a phrase."URLs and code" is too broad."web links and html" is better."http links and class attributes" is more specific."web addresses and code structures"Considering the `Yii::` pattern, it's not just HTML, but also executable code markers.Let's re-examine the `TOP_POSITIVE_LOGITS`: `Int`, `coded`, `/devices`, `spiel`, `ministries`, `stell`, `enso`, `)+`, `pawn`, `mighty`. None of these directly scream "URL" or "HTML attribute". But `/devices` could lean towards web services. `coded` is understandable. `spiel` is a bit random. `ministries` and `pawn` also seem random.However, the `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`, combined with `TOP_ACTIVATING_TEXTS`, are very strong.* `https://` is dominant.* `class="` is present.* `Yii` is present.The pattern is distinctly **web-related**.How about focusing on the most frequent element (`https` followed by `://`) and the structural elements (`class="`, `Yii::`)?"http/https links and code""web addresses and code elements""URLs, http, https, class, Yii" - too much listing."web links and code fragments"Let's try to be very specific about what it's detecting. It's detecting the *start* of URLs and the *start* of HTML attributes. It's also detecting `Yii::`.The combination of URL prefixes (`http`, `https`) *and* code attributes (`class=`) *and* framework calls (`Yii::`) strongly suggests it's identifying parts of web code."http, https, class attributes, Yii" - Still listing too much.The rule says "Avoid simply listing all the tokens. Instead, try to find patterns."The pattern is: elements that initiate web addresses or web code structures.Let's simplify further.- URLs

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    decoderesearch/olmo-3-saes/olmo-3-1025-7b
    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
    ady
    -0.09
    lys
    -0.09
    etten
    -0.08
    elige
    -0.07
    bil
    -0.07
     Bis
    -0.07
     Gross
    -0.07
    åľº
    -0.07
    489
    -0.07
     "\<
    -0.07
    POSITIVE LOGITS
    (Int
    0.09
    coded
    0.09
    /devices
    0.09
    spiel
    0.09
     ministries
    0.09
    stell
    0.08
    enso
    0.08
    )+↵
    0.08
    pawn
    0.08
     mighty
    0.08
    Activations Density 0.793%

    No Known Activations