© 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. Gemma-3-27B-IT
    3. 37-GEMMASCOPE-2-TRANSCODER-262K
    4. 203710
    Prev
    Next
    INDEX
    Explanations

    * **com** -> você (Portuguese "with" -> "you")* **trabalho** -> em (Portuguese "work" -> "in")* **ser** -> tú (Portuguese "to be" -> Spanish "you")* **ende** -> um (Portuguese "ende" -> "a/an")* **cooper** -> ativo (Portuguese "cooper" -> "active")* **Particip** -> ativa (Portuguese "Particip" -> "active")* **colaboración** -> de (Spanish "collaboration" -> "of")* **confidence** -> Brazilian (English "confidence" -> English "Brazilian") The tokens following these words often relate to actions, people, or contexts. "ativo" and "ativa" (active) are notable. "você", "tú", "you" suggest addressing someone.3. **TOP_POSITIVE_LOGITS**: * Exercises * exercised * تمارين (Arabic for "exercises") * ద (Telugu character, likely part of a word) * Dare * exercises * 执 (Chinese character, meaning "hold" or "grasp"; could be part of a word) * Pact * exercises * dare This list contains words related to "exercises", "dare", and "pact". It's highly multilingual (English, Arabic, Telugu, Chinese).4. **TOP_ACTIVATING_TEXTS**: * "celebrar este momento especial com você!" (Portuguese: "celebrate this special moment with you!") - contains "celebrar" (celebrate), "momento" (moment), "especial" (special), "com" (with), "você" (you). * "Estimula a criatividade, o trabalho em equipe, a diversão e a interação familiar." (Portuguese: "It stimulates creativity, teamwork, fun, and family interaction.") - contains "trabalho em equipe" (teamwork). * "Eres increíble y me haces muy feliz. Gracias por ser tú." (Spanish: "You are incredible and make me very happy. Thanks for being you.") - contains "ser tú" (being you). * "Defende um Estado forte e atuante na economia..." (Portuguese: "Defends a strong and active State in the economy...") - contains "atuante" (active). * "S (España): Algunas cajas rurales tienen un enfoque más ético y cooperativo." (Spanish: "(Spain): Some rural banks have a more ethical and cooperative approach.") - contains "cooperativo" (cooperative). * "Democracia Participativa: Ir más allá de la democracia representativa, permitiendo a los ciudadanos participar activamente en la toma de..." (Spanish: "Participatory Democracy: Go beyond representative democracy, allowing citizens to actively participate in the decision-making of...") - contains "Participativa" (Participatory), "participar" (participate), "activamente" (actively). * "El proyecto se basa en la colaboración de voluntarios que contribuyen con datos, entrenamiento, evaluación y desarrollo del modelo." (Spanish: "The project is based on the collaboration of volunteers who contribute data, training, evaluation, and model development.") - contains "colaboración" (collaboration). * "CPQD ... Foundation is a Brazilian leading center for research and development in telecommunications and information technology." - contains "Brazilian", "development". * "take his business to the next level."**Defining the Neuron's Scope**I am currently synthesizing the information from the provided lists to pinpoint the neuron's specific function. The goal is to create a concise descriptor that captures its essence.* **MAX_ACTIVATING_TOKENS:** `com`, `trabalho`, `ser`, `ende`, `cooper`, `Particip`, `colaboración`, `a`, `confidence` * These keywords relate to concepts like working together, being, participation, collaboration, and trust. Several are in Portuguese (`trabalho`, `ser`) and Spanish (`colaboración`).* **TOKENS_AFTER_MAX_ACTIVATING_TOKEN:** `você`, `em`, `tú`, `um`, `ativo`, `ativa`, `de`, `Brazilian`, `and` * This list suggests contexts or continuations like "you" (`você`, `tú`), locations/states (`em`, `de`), and descriptions (`ativo`, `ativa`, `Brazilian`). The combination of Romance languages and English is notable.* **TOP_POSITIVE_LOGITS:** `Exercises`, `exercised`, `تمارين`, `Dare`, `Pact` * This list highlights themes of activity, challenge ("Dare"), and agreement/commitment ("Pact"). The multilingual nature (English, Arabic) is clear.* **TOP_ACTIVATING_TEXTS:** * Texts mention celebrating with "you," "teamwork," "being you," "active State," "cooperative approach," "Participatory Democracy," "actively participate," "collaboration of volunteers," "Brazilian leading center for research and development," and "take his business to the next level."**Pattern Synthesis:**1. **Collaboration & Participation:** Words like `trabalho`, `cooper`, `Particip`, `colaboración`, `teamwork`, `participate`, `collaborative` appear frequently.2. **Action & State:** Terms like `ser`, `ativo`, `ativa`, `active`, `being`, `Exercises`, `Dare` suggest actions, states of being, or challenges.3. **Context/Recipient/Descriptors:** Tokens like `você`, `tú`, `you`, `em`, `um`, `de`, `a`, `Brazilian`, `and` provide context or link to participants/locations.The neuron seems to be capturing instances where individuals or groups are actively involved in a shared endeavor, project, or a state of having "confidence" or engagement. The presence of "exercises" and "dare" also hints at challenge or practice related to these collaborative efforts. The multilingual aspect suggests it's not tied to a single language but rather the underlying concepts.Combining these, a strong theme is "active participation" or "engaged collaboration." The "confidence" and "business" context from `TOP_ACTIVATING_TEXTS` also point towards professional or project-oriented collaborations.Let's try to be more specific based on the combination of `cooper`, `Particip`, `colaboración`, `trabalho em equipe` with `ativo`/`ativa` and `confidence`/`business`.Final proposed explanation: **active collaboration and confidence** (4 words)This fits the 3-20 word constraint.It captures the core concepts from all lists.It avoids forbidden phrases.active collaboration and confidence

    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_37_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
     குண
    0.40
     flying
    0.39
    ezi
    0.39
     Ezra
    0.39
     overflow
    0.39
    CTED
    0.38
    zu
    0.38
     ez
    0.38
     mastic
    0.37
    xmin
    0.37
    POSITIVE LOGITS
     Exercises
    0.43
     exercised
    0.39
     تمارين
    0.38
     ద
    0.38
    Dare
    0.36
     exercises
    0.36
    执
    0.35
     Pact
    0.35
    exercises
    0.35
     dare
    0.35
    Activations Density 0.014%

    No Known Activations