© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    APIAssistant AxisNEWCircuit TracerNEWSteerSAE EvalsExports Community BlogPrivacy & TermsContact
    1. Home
    3. Joseph Bloom · Feature Splitting for GPT2-Small
    5. GPT2-Small
    7. Residual Feature Splitting 12288
    9. 8-RES_FS12288-JB
    11. 664
    Prev
    Next
    INDEX
    Explanations

    negative statements or contradictions

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

    uncertainties and predictions regarding future events or outcomes.

    oai_token-act-pair · gpt-4o-miniTriggered by @supagao7

    the word “not” marking a negation (especially in the phrase “It’s not …”).

    oai_token-act-pair · o4-miniTriggered by @supagao7
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GPT2-SMALL @ 8-res_fs12288-jb
    Configuration
    jbloom/GPT2-Small-Feature-Splitting-Experiment-Layer-8/blocks.8.hook_resid_pre_12288
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    Skylion007/openwebtext
    Features
    12,288
    Data Type
    torch.float32
    Hook Point
    blocks.8.hook_resid_pre
    Architecture
    standard
    Context Size
    128
    Dataset
    Skylion007/openwebtext
    Hook Point Layer
    8
    Activation Function
    relu
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Neuron Alignment
    Index
    Value
    % of L₁
    94
    +0.11
    0.5%
    200
    +0.10
    0.5%
    674
    +0.10
    0.5%
    Correlated Neurons
    Index
    P. Corr.
    Cos Sim.
    200
    +0.11
    0.07
    437
    +0.10
    0.06
    94
    +0.10
    0.06
    Negative Logits
    çļ
    -0.79
    oided
    -0.72
    former
    -0.72
    çīĪ
    -0.67
    ãĤ¼
    -0.64
    assetsadobe
    -0.64
    ãģ®ç
    -0.64
    WAY
    -0.63
    éĥ
    -0.62
    ourses
    -0.62
    POSITIVE LOGITS
     necessarily
    1.39
    icable
    1.18
     uncommon
    1.11
    eworthy
    1.04
     exactly
    1.02
     entirely
    1.00
    orious
    0.99
     easy
    0.92
    icably
    0.91
    epad
    0.91
    Activations Density 0.183%

    No Known Activations