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

    Neuronpedia

    APICircuit TracerNEWSteerSAE EvalsExportsSlackBlogPrivacy & TermsContact
    © Neuronpedia 2025
    Privacy & TermsBlogGitHubSlackTwitterContact
    1. Home
    2. Google DeepMind · Exploring Gemma 2 with Gemma Scope
    3. Gemma-2-2B
    4. Residual Stream - 16k
    5. 20-GEMMASCOPE-RES-16K
    6. 2937
    Prev
    Next
    INDEX
    Explanations

    references to individual identity and personal experiences.

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

    gender-related discussions and possibly LGBTQ+ content.

    oai_token-act-pair · gpt-3.5-turboTriggered by @bot

    The main thing this neuron does is find the beginning of first-person statements or observations, particularly those starting with "I" or "And I" or similar phrases like "it was" or "it's". It seems to activate strongly for the start of personal opinions, reflections, or

    oai_token-act-pair · claude-3-5-sonnet-20240620Triggered by @bot

    "I" followed by a verb

    np_acts-logits-general · gemini-2.0-flash

    I followed by verbs

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GEMMA-2-2B @ 20-gemmascope-res-16k
    Configuration
    google/gemma-scope-2b-pt-res/layer_20/width_16k/average_l0_71
    Prompts (Dashboard)
    36,864 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    Features
    16,384
    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
     however
    -0.95
    however
    -0.90
    évaluateur
    -0.79
     echter
    -0.77
    However
    -0.76
     jednak
    -0.70
     však
    -0.66
     However
    -0.66
     porém
    -0.65
     però
    -0.64
    POSITIVE LOGITS
     correspondingly
    1.11
     accordingly
    1.10
     consequently
    1.05
    何より
    1.05
     consequent
    0.91
     derfor
    0.87
     frankly
    0.87
     certainly
    0.85
     dermed
    0.85
     соответственно
    0.84
    Activations Density 0.235%

    No Known Activations