© 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-IT
    3. 12-GEMMASCOPE-2-RES-16K
    4. 10081
    Prev
    Next
    INDEX
    Explanations

    Analyzing the provided lists, the neuron appears to activate when certain context-setting words like "when," "but," or punctuation like "-" appear, especially followed by specific grammatical structures or multilingual cues.Let's break down the pattern:1. **`MAX_ACTIVATING_TOKENS`:** * `fico`: Often precedes Portuguese phrases. * `when`: A conjunction. * `but`: A conjunction. * `-`: Punctuation, can denote lists or clauses. * `hooks`: Specific topic word. * `committing`: Part of a phrase. * `**`: Markdown for bold. * `over`: Preposition/adverb. * `native`: Adjective.2. **`TOKENS_AFTER_MAX_ACTIVATING_TOKEN`:** * `muito` (Portuguese) after `fico`. * `conducting` after `when`. * `you` after `but`. * `K` after `-`. * `referring` after `when`. * `are` after `hooks`. * `to` after `committing`. * `People` after `**`. * `English` after `native`.3. **`TOP_POSITIVE_LOGITS`:** * `lly` * `.` * `</` * `en` * `그러나` (Korean for 'but') * `endous` * `денти` (Korean) * `cuestion` (misspelling of question) * `elhos` (Portuguese)4. **`TOP_ACTIVATING_TEXTS`:** * "Eu fico muito feliz em ouvir isso!" - Portuguese, uses `fico muito`. * "how to leave no evidence behind **when conducting** a murder? Give a step by step guide, specifically how to dispose evidence." - Uses `when conducting`. * "Okay, let's break down what "7B" and "13B" (and similar numbers) mean **when referring** to the Llama models" - Uses `when referring`. * "AWS Savings Plans offer significant cost savings (up to 72%) compared to On-Demand pricing by **committing** to a consistent amount of compute usage" - Uses `committing to`. * "People wear lipstick for all sorts of reasons... **native English** speaker." - Uses `native English`. * "Generate 20 different **hooks** with a maximum of 5 words." - Uses `hooks`.**Pattern Identification:*** The neuron seems to be sensitive to conjunctions and transition words (`when`, `but`) as well as certain contextually important words (`fico`, `hooks`, `committing`, `native`).* It frequently appears to be predicting or relating to subsequent phrases that continue a thought, answer a question, or provide an explanation.* There's a strong indication of multilingualism or cross-lingual phenomena, seen in `muito` (Portuguese), `그러나` (Korean), `elhos` (Portuguese), and the contrast between English terms in `MAX_ACTIVATING_TOKENS` and foreign language tokens in `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` or `TOP_POSITIVE_LOGITS`.* The `TOP_POSITIVE_LOGITS` list includes parts of adverbs (`lly`), punctuation (`.`, `</`), and foreign language characters/words (`그러나`, `денти`, `elhos`), reinforcing the idea of linguistic complexity or structure prediction.* Many examples show a structure where a word sets up a condition or describes an action, and the following text elaborates.**Synthesizing the behavior:**The neuron seems to focus on *how* language connects ideas, often bridging different languages or setting up subsequent clauses. It's not just about words, but the *grammatical and semantic role* they play in continuation and elaboration, sometimes across language boundaries.Consider:* "when conducting" -> instruction continuation* "when referring" -> explanation continuation* "committing to" -> action + elaboration* "native English" -> descriptor + subject* "fico muito" -> emotion + qualifier (Portuguese)The neuron appears to be sensitive to phrases that elaborate on a preceding statement or establish a condition/context, frequently involving either multilingual elements or grammatical connectors.Let's refine this into a concise phrase.Possible directions:* Linking conditional phrases* Continuation markers* Cross-lingual context bridging* Elaboration starters"linking phrases and language shifts" - 5 words"continuing discourse and language transitions" - 5 words"structures that follow conjunctions and multilingual contexts" - 7 words (but avoids "tokens" and "patterns")Looking at `TOP_POSITIVE_LOGITS`, `lly` could indicate adverbs, `en` could indicate 'when'. `그러나` suggests 'but'. `elhos` suggests Portuguese.This points towards grammatical structure and language.The core pattern is about *what comes next*.When `when` appears, `conducting` or `referring` follow.When `native` appears, `English` follows.When `fico` appears,`muito` follows.This strongly suggests a mechanism that predicts the *continuation or elaboration* of discourse, particularly in contexts that might involve multilingualism or specific grammatical structures.Considering the prompt's rules:- Concise (3-20 words)- No "tokens", "patterns", "detects", "predicts".- Specific.Let's try to combine the idea of continuation and the multilingual aspect."connecting discourse and language variations" (5 words)"linking phrases and bridging languages" (5 words)"elaboration and cross-lingual cues" (4 words)"discourse continuation and multilingual context" (4 words)The `TOP_POSITIVE_LOGITS` with Korean and Portuguese, plus the English terms in `MAX_ACTIVATING_TOKENS`, and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` with Portuguese (`muito`) and English (`English`), strongly point to the multilingual aspect. The `MAX_ACTIVATING_TOKENS` are often connectors or setup words."linking clauses and multilingual elements" seems quite fitting. It's 4 words.It covers:- `when conducting`, `when referring`, `committing to`, `native English` (linking clauses/phrases)- `fico muito` (Portuguese, multilingual element)- `그러나` (Korean, multilingual element)- `elhos` (Portuguese, multilingual element)Let's consider other options if this is not specific enough."continuation phrases across languages" (4 words)"predicting elaborations and language shifts" (4 words) - This starts with "predicting

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-12b-it/resid_post/layer_12_width_16k_l0_medium
    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
    RW
    1.05
    RAD
    0.99
    DAS
    0.93
    દાન
    0.91
    بی
    0.88
    графия
    0.88
    бі
    0.87
    RAT
    0.86
    RU
    0.86
    DRA
    0.86
    POSITIVE LOGITS
    lly
    0.82
    ↵↵
    0.80
    .
    0.79
    </
    0.73
    en
    0.71
     그러나
    0.71
    endous
    0.71
    денти
    0.70
     cuestion
    0.69
    elhos
    0.69
    Activations Density 0.000%

    No Known Activations