© 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-3-12B
    3. 24-GEMMASCOPE-2-RES-16K
    4. 13242
    Prev
    Next
    INDEX
    Explanations

    **Explaining Neuron Behavior**I'm currently focused on articulating the neuron's actions by correlating `TOP_POSITIVE_LOGITS`, `TOP_ACTIVATING_TEXTS`, `MAX_ACTIVATING_TOKENS`, and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` into a succinct descriptive phrase.Let's analyze the provided data:* **MAX_ACTIVATING_TOKENS**: `know`, `Builder`, `k`, `0`, `CE`, `/`, `[`* **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `(`, `P`, `setting`, `Add`, `Date`, `±`, `s`, `**`, `Number`, `Age`Looking at the `TOP_ACTIVATING_TEXTS` and `MAX_ACTIVATING_TOKENS`:1. "...something you **know** (PIN)..." - This is a direct match for `know`.2. "Chart**Builder** AddLegend()" - This is a direct match for `Builder`.3. "...11th century **CE**..." - This is a direct match for `CE` with the token `/` immediately following.4. "...numbers of the form `6k ± 1`." - This matches `k` and `±`.The `TOP_POSITIVE_LOGITS` include suffixes like `able`, `ized`, `coll`, `isan`. This suggests the neuron might be related to word formation or specific types of words. The presence of `Walk` is interesting but might be less dominant than the suffix patterns or specific tokens.Let's try to find a common theme or pattern:- "know" + (PIN) -> password/authentication- "Builder" + Add -> object creation, coding syntax (like builder pattern)- "CE" + / or date-like context (Number, Date) -> historical periods, years- "k" +/- 1 -> mathematical sequences, primes.The neuron seems to be sensitive to specific words (`know`, `Builder`, `CE`) and what follows them. It also seems to activate on numerical/alphanumeric sequences.Considering the `TOP_POSITIVE_LOGITS` which are mostly suffixes (`able`, `ized`, `coll`, `isan`), and the specific tokens appearing. The combination of specific tokens like "know", "Builder", "CE" and the prevalence of word endings suggests a function related to identifying specific grammatical structures or specific types of informational content.Let's re-examine the `TOP_ACTIVATING_TEXTS`:- MFA: "something you **know** (PIN)" - "know" is followed by a parenthesized explanation.- Chart**Builder** AddLegend() - "Builder" is followed by a method call.- "11th century **CE**" - "CE" is followed by context about historical periods.The neuron is detecting specific keywords or codes and then the context that follows them, which can be varied (like a parenthesized explanation, a method name, or a time period indicator). The logits also suggest an interest in word endings which often denote states or actions (`-able`, `-ized`, `-er` from Builder).The phrase "know", "Builder", "CE" seems to be common nodes of activation.Let's look at the actual tokens again.MAX_ACTIVATING_TOKENS: `know`, `Builder`, `k`, `0`, `CE`, `/`, `[`TOKENS_AFTER_MAX_ACTIVATING_TOKEN: `(`, `P`, `setting`, `Add`, `Date`, `±`, `s`, `**`, `Number`, `Age`The pattern appears to be identifying certain keywords (`know`, `Builder`, `CE`) and then recognizing what kind of information follows them, like explanations, code structures, or temporal markers.Let's try to synthesize this. The neuron is not just identifying one thing, but a relationship or a type of construct.Consider "know (PIN)", "Builder Add", "CE (year context)".It seems to pick up key terms that introduce specific definitions, actions, or time references.Could it be related to specific input fields or parameters?- Know -> Authentication parameter- Builder -> Object declaration/construction- CE -> Date/Period markerGiven the explanation should be concise and specific:"know (PIN)", "Builder Add", "CE /", "6k ± 1"The commonality is specific entities followed by explanatory or structural elements.The neuron seems to pick up on specific keywords followed by qualifying information.Let's look at the positive logits again: `са`, `ina`, `able`, `coll`, `isan`, `ди`, `Walk`, `ized`, `к`, `ا`. Many are suffixes indicating properties or states.The combination of specific tokens like `know`, `Builder`, `CE`, `k` and the `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` like `(`, `Add`, `±`, `Number` points to recognizing structured information or specific types of terms."know" -> authentication context"Builder" -> code construction context"CE" -> time/era context"k ±" -> mathematical sequencesThis neuron seems to be recognizing specific keywords that are often followed by parameters, definitions, or specific contextual information. The suffixes in the logits might reinforce this by suggesting it's looking for words that *define* or *modify*.Let's try to find a unifying concept. It's about recognizing specific labels and their associated data or function.What about focusing on the "knowing" or "identifying" aspect?"know", "Builder", "CE", "k" are specific identifiers.The tokens after provide the context or function.How about "Keywords and their context"? This is 4 words."Specific terms and follow-up details" - 5 words."Keyword recognition with specific context" - 5 words.Let's check if any single word or very short phrase could capture this.The `TOP_POSITIVE_LOGITS` are interesting. `able`, `ized`, `coll`, `isan` are often found in technical terms or descriptions.The phrase should be specific.Consider `know` followed by `(...)`, `Builder` followed by `Add`, `CE` followed by `/` (context separator or a year), `k` followed by `±`. These are all forms of `SPECIFIC_TERM` + `QUALIFIER`.The word "Builder" implies creating something. "know" implies information. "CE" implies a time.The pattern is recognizing these specific entities.Let's consider the most prominent parts:`know` + `(something)``Builder` + `Add``CE` + `/` or context`k` + `±`It's about recognizing specific identifiers and their associated structures or modifiers.The neuron might be identifying specific technical terms or conceptual labels.Let's look for a pattern that covers `know (something)`, `Builder Add`, `11th century CE

    np_acts-logits-general · gemini-2.5-flash-lite

    The neuron strongly activates on numeric or quantifying tokens—numbers, dates, time units, counts, monetary figures, etc.

    oai_token-act-pair · o4-miniTriggered by @jyhe0408
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-12b-pt/resid_post/layer_24_width_16k_l0_medium
    Prompts (Dashboard)
    392,802 prompts, 256 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     drawSprites
    0.86
    ోతి
    0.86
    ứng
    0.84
    ㅑ
    0.82
    raya
    0.81
    믈
    0.81
    ים
    0.80
     İstifadə
    0.79
    ེད་
    0.77
    ệc
    0.76
    POSITIVE LOGITS
    са
    0.88
    ina
    0.78
    Walk
    0.70
    ized
    0.69
    т
    0.69
    isan
    0.68
     coll
    0.67
    ا
    0.67
    ba
    0.67
    ди
    0.66
    Activations Density 0.000%

    No Known Activations