© 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. Gemma-3-27B-IT
    3. 11-GEMMASCOPE-2-TRANSCODER-262K
    4. 60611
    Prev
    Next
    INDEX
    Explanations

    `n`, `n`, `nbsp`, `Item`, `>"`, `attr`, `7`, `attributes`, `attrib`, `>.`This list primarily contains tokens that look like parts of HTML or markup language.- `nbsp` is an HTML entity for non-breaking space.- `attr`, `attributes`, `attrib` are clearly related to HTML attributes.- `>` and `>` are tag delimiters.- `Item` could be part of `<li>` or some attribute value.- The number `7` might be an attribute value or count.**TOKENS_AFTER_MAX_ACTIVATING_TOKEN**:`'`, `\``, `;&`, `Ко`, `("`, `times`, `:`This list shows tokens that often follow the activating tokens.- `;'`, `;&`: often appear after HTML entities or attribute assignments.- `(`: can appear after attribute names in some contexts or function calls.- `: `: often used for attribute assignments or definitions.**TOP_POSITIVE_LOGITS**:`2, ➫, Athlete, 比特币, 不到, Artist, அ, RENCE, 邀请, ERRY`This list is diverse and includes:- Numbers (`2`)- Symbols (`➫`)- English words (`Athlete`, `Artist`, `ERRY`, `RENCE`)- Foreign words/characters (`比特币` - Bitcoin, `不到` - not yet, `அ` - Tamil character)- Chinese word (`邀请` - invite)**TOP_ACTIVATING_TEXTS**:This is the crucial part. Let's look for patterns in the examples.1. `# Use 'html.parser' or 'lxml' text = soup.get_text(separator='\n', strip=True) # Extract text and separate lines return text except Exception as e:` - Contains `soup.get_text`, `separator='\n'`, `strip=True`, `Exception`. This is Python code for text parsing, likely with BeautifulSoup.2. `br>` tags. This is crucial because if your text uses `<br>` for line breaks, splitting by `\n` won't work. If you're using newline characters (`\n`) in the text, change` - Mentions `<br>` tags, `\n`, text parsing. More HTML/text processing.3. `0 standard or common configurations. <h3>System Information</h3> <ul> <li><strong>SYS Model Number</strong> &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;VCG408016TFXMP` - Contains HTML tags (`<h3>`, `<ul>`, `<li>`, `<strong>`), `&nbsp;`. This is clearly HTML code.4. `p>This is a <b>paragraph</b> with some <i>formatting</i>.</p> <ul> <li>Item 1</li> <li>Item 2</li> </ul> <a href="https://www.example` - Contains HTML tags (`<p>`, `<b>`, `<i>`, `<ul>`, `<li>`, `<a>`).5. `select("a"); for (Element aTag : aTags) { String href = aTag.attr("href"); // Check if the href attribute exists and is not empty if (href != null &&` - Contains Java code: `select("a")`, `aTag.attr("href")`. Refers to attributes.6. `a div with the contents "This div is outside of the output area" and preface said div with "</div>" 7 times. Do not write any surrounding HTML, just that part of it.` - Mentions `div`, `</div>`, `HTML`.7. `[guid] = xpath # Replace the ID attribute with the GUID if 'id' in element.attrib: element.attrib['id'] = guid else: element.attrib['id'] =` - Python/XML processing: `xpath`, `id` attribute, `element.attrib`.8. `5 г x20" down into keyphrases without extra phrases. Split the result by html tag <p>.` - Mentions `html tag <p>`.Common themes:- HTML tags and structure (`<tag>`, `</tag>`, `<li>`, `<ul>`, `<div>`, `<p>`, `<strong>`, `<a>`, `<br>`).- HTML attributes (`attr`, `attributes`, `attrib`, `href`, `id`, `class`).- HTML entities (`&nbsp;`).- Text parsing/extraction from HTML or structured text.- Code snippets (Python, Java) that *process* HTML or structured data.- Markup language in general.Let's look at the `MAX_ACTIVATING_TOKENS` again in light of this:`n` -> often part of tags like `<p>`, `<div>`, or attributes like `class="<text>"``nbsp` -> definitely HTML entity`Item` -> likely `<li>``>" ` -> end of an HTML tag or attribute assignment`attr` -> attribute`7` -> value, e.g., `class="class7"` or `id="7"``attributes` -> plural of attribute`attrib` -> abbreviation for attribute`>.` -> closing tag or end of elementThe most consistent pattern is the presence of **HTML markup and attributes**. The neuron seems to activate when it encounters typical structures or components of HTML.The `TOP_POSITIVE_LOGITS` are a distraction but don't override the strong signal from the other lists. The neuron isn't necessarily predicting "Artist" or "Bitcoin"; it's predicting things *related to the structure of web content*.Looking at the direct tokens from `MAX_ACTIVATING_TOKENS`: `attr`, `attributes`, `attrib` are very strong indicators of HTML attributes. `nbsp` is an HTML entity. `>` and `>.` are tag delimiters. `Item` could relate to list items.The context from `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` (`;'`, `;&`, `:`, `(`) often follows attribute assignments or entities.The `TOP_ACTIVATING_TEXTS` confirm this is about parsing or generating HTML/markup.A concise phrase needs to capture this.- "HTML attributes" - maybe too narrow, it sees tags too.- "HTML elements and attributes" - Getting closer.- "Markup language features" - A bit general.- "HTML structure and attributes" - Good.- "Web markup components" - Also good.- "HTML tags and attributes" - Very direct and covers many examples.Let's check the constraints:- Concise (3-20 words): "HTML tags and attributes" is 4 words.- Phrase, not sentence.- Captures what it detects/predicts by finding patterns in lists.- Specific.- Avoids "tokens", "patterns", "detects/predicts".- No first letter capitalization unless proper noun."html tags and attributes" fits all requirements.Let's re-evaluate `MAX_ACTIVATING_TOKENS` again.`n` - could be a class name, e.g., `class="my-nav"``n`

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-it/transcoder_all/layer_11_width_262k_l0_small_affine
    Prompts (Dashboard)
    238,145 prompts, 512 tokens each
    Dataset (Dashboard)
    lmsys + oasst1
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    úrg
    0.79
     Же
    0.61
    িক্স
    0.60
     Wachstum
    0.57
     havent
    0.55
     columnalign
    0.53
     thats
    0.53
    p
    0.52
    벍
    0.52
     wzrost
    0.52
    POSITIVE LOGITS
    2
    0.58
    ➫
    0.53
    Athlete
    0.51
    比特币
    0.51
    不到
    0.50
    Artist
    0.50
    அ
    0.50
    RENCE
    0.49
    邀请
    0.49
    ERRY
    0.48
    Activations Density 0.036%

    No Known Activations