© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    Natural Language
    Autoencoders
    NEW
    Assistant AxisNEWCircuit TracerUPDATESteerSAE EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Qwen3.5-4B
    3. 15-RES-MATRYOSHKA-65K
    4. 19706
    Prev
    Next
    INDEX
    Explanations

    `argument. As a result, particle verbs are argued to` * `exRecognizer", "intents": [ { "intent": "CreateMeetingIntent", "pattern": "(?i)create meeting" }, { "intent": "HelpIntent` * `separations and full utilization of the enhanced 2D peak capacity is selective exploitation of the sample attributes, described as sample dimensionality. In order to take sample dimensionality into account prior to optimization of a 2D separation, a new concept` * `features. However, the recent studies have proved that weak object-based theory because the features from the same dimension (e.g., two colors) cannot` * `consequences. A discrete choice experiment (DCE) quantified value-based preferences for alternative cannabis policies described by the five key attributes legal status, health harms, criminal justice` * `valid if [[ $1 == "--nightly" ]]; then echo "Building for nightly release" NIGHTLY=true # Get Formatted Date BUILD_DATE="$(date +%m_%d_%Y)" elif` * `stored in a variable a <- expression(10 + x + y) I want to use substitute to fill the expression with x = 2 substitute(a, list(x=2)) But this returns a and a evaluates` * `on a set of rules. Typically, an application client that initiates a full-text search has to maintain the filter. That is` * `definition.DocumentGraphValidator; import com.yahoo.searchdefinition.Search; import com.yahoo.searchdefinition.SearchBuilder; import com.yahoo.searchdefinition.document.SDDocumentType; import com.yahoo.searchdefinition.parser.ParseException; import org.junit.Rule;`Let's look for patterns:**MAX_ACTIVATING_TOKENS**: `dimension`, `Ground`, `dimension`, `.,`, `attributes`, `NIGHT`, `the`, `filter`, `com`**TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `categories`, `argument`, `"`, `ality`, `two`, `legal`, `LY`, `expression`, `.`, `.yahoo`Observations:* `dimension` appears twice in MAX_ACTIVATING_TOKENS. `attributes` appears there too.* `dimension` is often followed by "categories" or "attributes" or "values" etc. in the text. Example: "dimension may be described as a categorical attribute", "sample attributes, described as sample dimensionality".* `Ground` is followed by `argument` (in the TEXT).* `NIGHT` is followed by `LY` (in TOKENS_AFTER) and likely relates to "nightly" (TEXT).* `filter` is followed by `.` or things related to programming/definitions.* `com` is followed by `.yahoo` (in TOKENS_AFTER).Let's try to find a common theme or category.The texts discuss:* Data dimensionality, categorical attributes, measures.* Abstract concepts like "Ground argument".* Programming/technical aspects: `exRecognizer`, `intent`, `pattern`, `expression`, `variable`, `filter`, `com`, `yahoo`.* Specific technical terms/proper nouns (like `NIGHTLY` or domain-specific terms in Chinese in `TOP_POSITIVE_LOGITS`).The neuron seems to be activated by technical or specific domain terms.The `MAX_ACTIVATING_TOKENS` list has a mix of general technical terms (`dimension`, `attributes`, `filter`) and some more specific ones (`NIGHT`, `Ground`, `com`).The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list also shows context: `categories`, `argument`, `expression`, `.yahoo`.Let's look at `dimension`, `attributes`, `filter`. These are generally related to data representation or programming structures.The presence of `NIGHTLY`, `expression`, `com`, `yahoo` suggests a technical context.Perhaps it's about defining or organizing something."dimension" -> "categories""Ground" -> "argument""NIGHT" -> "LY" (like nightly)"filter" -> ".""com" -> ".yahoo"These look like components or contexts within technical definitions or code.The phrase "Processing Neuron Logic" is not good."Technical specification components" - a bit long."technical context identifiers" - good, but maybe too generic.Let's check the TOP_POSITIVE_LOGITS again: `ä¸ŃçļĦæķ°æį®`, `tac`, `ettes`, `ures`, `ultan`, `æIJľç´¢èµĦæĸĻ`, `被åĪ«äºº`, `çļĦ人æķ°`, `(!((`, `tac`These look like non-English words, possibly code or technical jargon from other languages. This really pushes the 'technical/specific' angle.The pattern seems to be related to technical terms, possibly identifiers, data structures, or specific keywords.`dimension`, `attributes`, `NIGHT`, `filter` are keywords in data analysis or programming.`Ground`, `argument` has a linguistic/computational linguistics feel.`com`, `yahoo` is specific to web/tech.Maybe it's about identifying specific *types* of terms or identifiers.Let's try to find a more exact match theme.`dimension` followed by `categories` or `attributes``Ground` followed by `argument``NIGHT` followed by `LY` (forming `nightly`)`filter` is a keyword.`com` followed by `.yahoo` potentially implying a domain name component.The common thread is *components* of technical definitions, data structures, or code."technical definition components" - 3 words."technical keywords and identifiers" - 4 words."specific technical identifiers" - 3 words.Consider the range of:- `dimension` (data)- `NIGHT` (builds)- `filter` (programming)- `com` (domains)- `Ground` (linguistics/semantics)This neuron is quite general in the *types* of technical terms it activates on. But they are all *specific*."specific technical terms" - 3 words."keywords and technical definitions" - 4 words.Let's look at the actual values again.`dimension` ... `categories``Ground` ... `argument``NIGHT` ... `LY` (e.g. nightly)`filter` ... (e.g. filter rules)`com` ... `.yahoo`It's about identifying specific *components* or *elements* within larger technical contexts.The tokens after MAX_ACTIVATING_TOKENS are often the *next word* in a phrase or code snippet.`dimension` -> `categories``Ground` -> `argument``NIGHT` -> `LY``filter` -> `.``com` -> `.yahoo`This looks like detecting specific keyword/identifier followed by a specific common suffix or subsequent token."specific keyword and suffix" - 4 words."technical component recognition" - 3 words.Let's try to be more direct with the *type* of thing.The top positive logits are quite alien, suggesting it's picking up on highly specific, perhaps non-standard, identifiers or terms.`ä¸ŃçļĦæķ°æį®` is a very specific string.Okay, let's re

    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
    untu
    -0.07
    å¾ĢæĿ¥
    -0.06
    asto
    -0.06
    бо
    -0.06
    主è¦ģ
    -0.06
    le
    -0.06
    éĽĨä¸Ń
    -0.05
    æĮ¯
    -0.05
    igas
    -0.05
    ÑĩаÑģ
    -0.05
    POSITIVE LOGITS
    ä¸ŃçļĦæķ°æį®
    0.07
    tac
    0.06
    ettes
    0.06
    ures
    0.06
    ultan
    0.06
    æIJľç´¢èµĦæĸĻ
    0.05
    被åĪ«äºº
    0.05
    çļĦ人æķ°
    0.05
     (!((
    0.05
     tac
    0.05
    Activations Density 0.004%

    No Known Activations