Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    APIAssistant AxisNEWCircuit TracerNEWSteerSAE EvalsExports Community BlogPrivacy & TermsContact
    © Neuronpedia 2025
    Privacy & TermsBlogGitHubSlackTwitterContact
    1. Home
    2. Google DeepMind · Exploring Gemma 2 with Gemma Scope
    3. Gemma-2-9B-IT
    4. Residual Stream - 131k
    5. 20-GEMMASCOPE-RES-131K
    6. 85683
    Prev
    Next
    INDEX
    Explanations

    syntax-related elements or symbols in code

    oai_token-act-pair · gpt-4o-miniTriggered by @bot

    code delimiters and assignmentsRationale:The `MAX_ACTIVATING_TOKENS` list contains symbols like `=`, `:`, `{`, `}`, `[`, `]`, which are common delimiters and assignment operators in programming code.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list includes `this`, `new`, `NULL`, `is`, `match`, which are frequently found in code contexts following assignments or definitions.The `TOP_ACTIVATING_TEXTS` are clearly code snippets from various programming languages.The `TOP_POSITIVE_LOGITS` list is a bit mixed, but `AssemblyCompany`, `transQ`, `piram` could be seen as specific identifiers or project names within code. The inclusion of foreign language characters might indicate handling of internationalized code or comments.Overall, the neuron seems to activate when it encounters code structures involving assignments and delimiters, likely identifying specific types of code blocks or variable assignments. The phrase "code delimiters and assignments" captures this core observation concisely. code delimiters and assignments

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GEMMA-2-9B-IT @ 20-gemmascope-res-131k
    Configuration
    google/gemma-scope-9b-it-res/layer_20/width_131k/average_l0_81
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    Features
    131,072
    Data Type
    float32
    Hook Name
    blocks.20.hook_resid_post
    Hook Layer
    20
    Architecture
    jumprelu
    Context Size
    1,024
    Dataset
    monology/pile-uncopyrighted
    Activation Function
    relu
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     arşivlendi
    -0.49
    DockStyle
    -0.48
     judicia
    -0.41
    WriteTagHelper
    -0.40
     arqu
    -0.39
    explique
    -0.37
     embl
    -0.35
    ClientRect
    -0.35
    AddTagHelper
    -0.34
    absl
    -0.33
    POSITIVE LOGITS
     betweenstory
    0.69
    
    0.54
     AssemblyCompany
    0.54
    Datuak
    0.49
    transQ
    0.47
     виправивши
    0.47
     صوتيه
    0.46
    جوايز
    0.45
    piram
    0.45
     etc
    0.44
    Activations Density 0.013%

    No Known Activations