© 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 EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-27B-IT
    3. 31-GEMMASCOPE-2-RES-262K
    4. 43127
    Prev
    Next
    INDEX
    Explanations

    modelthoughtThe user wants me to explain the neuron's behavior concisely (3-20 words) by finding a pattern in the provided lists.**MAX_ACTIVATING_TOKENS**:- Thought (multiple times)- inkle- Leo- augh- Finn**TOKENS_AFTER_MAX_ACTIVATING_TOKEN**:- : (multiple times)- :** (multiple times)- >- AndObservation: The `MAX_ACTIVATING_TOKENS` seem to be internal thoughts, names, or prefixes to potentially represent spoken dialogue or structured thought processes within text. The tokens following them (`:`, `:**`) strongly suggest they are being used as separators, labels, or indicators before a continuation. 'Thought' is explicitly a marker for internal thinking. 'Leo' and 'Finn' are names that might preface dialogue or character actions.**TOP_POSITIVE_LOGITS**:- wäre (German for 'would be')- AttributeError (Python error)- workmanship- Millionen (German for 'millions')- 设计 (Chinese for 'design')- ጥ (Ethiopic character)- 勹 (Chinese radical)- Millions- Promised- Quelle (German for 'source')Observation: This list is diverse, including non-English words, programming terms, and general English words. It doesn't provide a clear semantic theme. The presence of `AttributeError` might suggest error handling or unexpected conditions. `Millionen`/`Millions` and `Promised` could relate to outcomes or values.**TOP_ACTIVATING_TEXTS**:- `Final Answer: the final answer to the original input question Question: how to earn passive income<end_of_turn> <start_of_turn>model Thought:` - Activates after a model's thought process.- `aby is in the back row, meticulously leveling his desk with a small bubble level.)** Ms. Periwinkle: …and so, the water evaporates, rises into the atmosphere, condenses into clouds, and then precipitates back down` - Contains names and dialogue markers.- `nBegin! \n\n \n Question: you are chatgpt Language model?<end_of_turn> <start_of_turn>model Thought:` - Activates after a model's thought process.- `. * **Visual:** Medium shot of Maya with the headset on. She looks deeply unimpressed. * **Leo:** "Details, details! It's perfectly safe! Just tell me what you're feeling *right now*` - Contains character names and dialogue.- `1. tools.end() </Action> <User> Hi! </User> <Thought> The user has greeted me. Let's respond with a friendly greeting. </Thought> <Action` - Contains `Thought` as an internal marker.- `Yelp.Search(), Spotify.play Begin! Human: which air fryer should i buy<end_of_turn> <start_of_turn>model Thought:` - Activates after a model's thought process.- `’s face, which is quite a distance from where he’s looking.)** **Mr. Featherstonehaugh:** (Stuttering) A-a-a remarkably… vibrant… tapestry, Mrs. Van Derlyn. internal thoughts and dialogue markers

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-it/resid_post/layer_31_width_262k_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
     televisión
    0.52
    government
    0.48
     छात्रा
    0.47
     mitotic
    0.47
    car
    0.46
     употреб
    0.46
     television
    0.46
    television
    0.46
    र्श
    0.45
     gobierno
    0.45
    POSITIVE LOGITS
     wäre
    0.52
     AttributeError
    0.47
     workmanship
    0.46
     Millionen
    0.44
     设计
    0.43
     ጥ
    0.42
    勹
    0.42
     Millions
    0.42
     Promised
    0.42
    Quelle
    0.41
    Activations Density 0.004%

    No Known Activations