© 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. Gemma-4-31B
    3. 30-RES-MATRYOSHKA-131K
    4. 119934
    Prev
    Next
    INDEX
    Explanations

    badge/%20compatible-swift%203.0-orange.svg)](http://swift.org) [![License](- Contains `compatible`.3. `monolog) [![Total Downloads](https://img.shields.io/packagist/dt/monolog/monolog.svg)](https://packagist.org/packages/monolog/monolog) [![Latest` - Contains `monolog`.4. `Q: How to use two conditions in "where" clause in XQuery I'm trying to extract only those` - Contains `Q:`. The `Q` token is present. The `:` token follows `Q`.5. `[Intracranial aneurysm and coronary ectasia]. Coronary aneurysms are uncommon and can exceptionally be associated with` - Contains `[Intracranial aneurysm...`. The `[` token is present. The `Int` token is part of the text following `[`.Looking at the patterns:- `&` is followed by `height` and `compatible` (in context of image badges/shields).- `Q` is followed by `:`.- `[` is followed by `Int` (part of a phrase).The most consistent pattern seems to be punctuation/special characters followed by specific words or tokens, often marking the start of a new element or query.The `MAX_ACTIVATING_TOKENS` are `&`, `0`, `/`, `Q`, `[`.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are `height`, `compatible`, `mon`, `:`, `Int`.The texts show:- `&months=6&height=3` -> `&` followed by `height`- `/%20compatible-swift%203.0-orange.svg` -> `/` followed by `compatible` (URL encoding for characters)- `monolog` -> no preceding special char from the list, but `monolog` is in the list.- `Q:` -> `Q` followed by `:`- `[Intracranial...` -> `[` followed by `Int`The pattern is specifically about certain punctuation or query-like characters followed by specific terms.Let's focus on `&`, `Q`, `[` from MAX_ACTIVATING_TOKENS and their corresponding followers.`&` → `height`, `compatible` (often in link/badge contexts)`Q` → `:` (in question format)`[` → `Int` (start of a title/phrase)This neuron seems to be activating on specific separators or markers (like `&`, `Q:`, `[`) followed by structured content or specific keywords."badge/link separators" is too long."Q: / & [" is just listing tokens."punctuation then keyword" is too general.Consider the `TOP_POSITIVE_LOGITS`: Many are non-English. However, 'abstinence' is English. This list is highly diverse.The prompt asks for what the neuron *detects or predicts by finding patterns in lists*.Let's look at the `MAX_ACTIVATING_TOKENS` as potential triggers.`&` -> often used in URLs or lists/arguments (e.g., `?param1=value1&param2=value2`)`Q` -> often starts a question or query. `Q:` is common.`[` -> often used for lists, arrays, or titles/sections. `[Intracranial aneurysm]`The tokens following are `height`, `compatible`, `mon`, `:`, `Int`.Combining `&` and `height`/`compatible` points to URL parameters or image badge construction.`Q:` points to queries.`[` and `Int` points to starting a specific section/phrase.The common theme seems to be structural markers followed by specific content.The most general, yet specific enough pattern found across multiple examples is the appearance of specific punctuation/symbols like `&`, `Q`, `[` followed by subsequent text, often defining structure or specific entities.Let's re-evaluate `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `height`, `compatible`, `mon`, `:`, `Int`.And `MAX_ACTIVATING_TOKENS`: `&`, `0`, `/`, `Q`, `[`.- `&` + `height` (from text 1)- `/` + `compatible` (from text 2, after URL encoding)- `Q` + `:` (from text 4)- `[` + `Int` (from text 5)It seems the neuron is sensitive to certain delimiters or initiating characters followed by specific words or structures.The explanation should be about what it *detects* or *predicts*.What about "start of structured links or queries"? This is 6 words."markup or query starters" - 4 words."special character followed by word" - 5 words.Let's look at the examples again:1. `&height`, `&months` from URLs.2. `/` ... `compatible` from a URL.3. `monolog` is a specific term, but `mon` is a follower token. It's like a standalone mention.4. `Q:` specific query format.5. `[` ... `Int` (part of `[Intracranial aneurysm]`)The neuron seems to detect specific structural elements or common prefixes for certain types of information: links/badges (using `&`, `/`), queries (`Q:`), and specific titled sections (`[` followed by notable words like `Int`)."special characters before specific words" - 5 words."structural markers before text" - 4 words.Let's consider the specific followers: `height`, `compatible`, `mon`, `:`, `Int`.And the specific precursors: `&`, `/`, `Q`, `[`.This feels like it's detecting structured data snippets, especially in code or markdown contexts."link/badge syntax""markdown and url structure""structured text delimiters"How about focusing on the `Q:` and `[` pattern?"query prefixes and lists" - 4 words.Let's try to be more literal about the combinations seen: `&height`, `&compatible`(implied in URL), `Q:`, `[Int`.The neuron detects a specific character followed by a specific word/structure."specific prefixes and continuations" - 4 words."query starters and link elements" - 5 words.The `TOP_POSITIVE_LOGITS` are very diverse and don't seem to offer a clear pattern for this specific explanation task.Let's focus on the most clear patterns:`&` -> `height``Q` -> `:``[` -> `Int`These are all about *following* certain tokens.The explanation could be about the tokens that follow `&`, `Q`, `[`, etc."tokens following specific punctuation" - 4 words.This is accurate. The tokens `height`, `compatible`, `mon`, `:`, `Int` are identified because they follow the tokens `&`, `/`, `Q`, `[`.Let's simplify this. The neuron is detecting these specific punctuation marks and then looking at the subsequent structure or word."symbols followed by specific words" - 4 words.This covers `&height`, `Q:`, `[

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    decoderesearch/gemma-4-saes/gemma-4-31b
    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
    你知道
    -0.07
     FORM
    -0.06
     vaulted
    -0.06
     most
    -0.06
    INCLUDE
    -0.06
     funzionalità
    -0.06
     reposition
    -0.06
    Make
    -0.06
    ※
    -0.06
     wakati
    -0.06
    POSITIVE LOGITS
    ૯
    0.06
     świę
    0.06
     기업
    0.06
    多かった
    0.06
     ieri
    0.06
     을
    0.06
     явления
    0.06
     abstinence
    0.05
    ऐ
    0.05
    uters
    0.05
    Activations Density 0.001%

    No Known Activations