© 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
    3. Google DeepMind · Exploring Gemma 2 with Gemma Scope
    5. Gemma-2-2B
    7. Residual Stream - 16k
    9. 25-GEMMASCOPE-RES-16K
    11. 14186
    Prev
    Next
    INDEX
    Explanations

    segments at the beginning of sequences.

    oai_token-act-pair · gpt-4oTriggered by @gulkoa

    references to legal or structured standards and their implications

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

    The neuron primarily detects the start-of-sentence marker, firing strongly on the <bos> token.

    oai_token-act-pair · o4-miniTriggered by @gulkoa

    the beginning of a document/sequence.

    oai_token-act-pair · gemini-1.5-proTriggered by @gulkoa

    the beginning of a sentence or text (high activation for <bos> token and other start-of-sentence constructions).

    oai_token-act-pair · claude-3-5-haiku-20241022Triggered by @gulkoa

    the start token (<bos>) indicating the beginning of text segments.

    oai_token-act-pair · deepseek-r1Triggered by @gulkoa

    the beginning of a sentence or text segment, as indicated by the high activation on the `<bos>` (beginning of sentence) token.

    oai_token-act-pair · deepseek-v3Triggered by @gulkoa

    the beginning of sentences or text segments, particularly focusing on the start of input tokens (like <bos>) and common sentence-starting words or phrases.

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

    the beginning of text, strongly activating for the <bos> token and initial words of documents.

    oai_token-act-pair · claude-3-7-sonnet-20250219Triggered by @gulkoa
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GEMMA-2-2B @ 25-gemmascope-res-16k
    Configuration
    google/gemma-scope-2b-pt-res/layer_25/width_16k/average_l0_116
    Prompts (Dashboard)
    36,864 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    Features
    16,384
    Data Type
    float32
    Hook Name
    blocks.25.hook_resid_post
    Hook Layer
    25
    Architecture
    jumprelu
    Context Size
    1,024
    Dataset
    monology/pile-uncopyrighted
    Activation Function
    relu
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    rungsseite
    -0.80
     nahilalakip
    -0.79
     estekak
    -0.77
    SourceChecksum
    -0.77
     ſei
    -0.77
     ſeine
    -0.77
     &___
    -0.77
     ſans
    -0.76
     deſſen
    -0.76
     ſelbſt
    -0.76
    POSITIVE LOGITS
     ögon
    -0.15
     stället
    -0.18
     varandra
    -0.19
     vandens
    -0.20
     kysy
    -0.20
     bēr
    -0.20
     vastaan
    -0.21
     méret
    -0.21
     tarko
    -0.21
     vägen
    -0.21
    Activations Density 34.919%

    No Known Activations