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

    Neuronpedia

    APISteerSAE EvalsBlog/PodcastNEWSlackPrivacy & TermsContact
    © Neuronpedia 2025
    Privacy & TermsBlog/PodcastGitHubSlackTwitterContact
    1. Home
    2. Joseph Bloom · Feature Splitting for GPT2-Small
    3. GPT2-Small
    4. Residual Feature Splitting 768
    5. 8-RES_FS768-JB
    6. 0
    Prev
    Next
    INDEX
    Explanations

    phrases related to international relations or policies

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

    information related to governmental policies and international affairs

    oai_token-act-pair · gpt-3.5-turboTriggered by @bot
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GPT2-SMALL @ 8-res_fs768-jb
    Configuration
    jbloom/GPT2-Small-Feature-Splitting-Experiment-Layer-8/blocks.8.hook_resid_pre_768
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    Skylion007/openwebtext
    Features
    768
    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
    •Test list
    "with"
    •"with" features in 768 and 1536
    •Canonical Fig 12

    No Comments

    Neuron Alignment
    Index
    Value
    % of L₁
    362
    +0.15
    0.7%
    640
    +0.13
    0.6%
    60
    +0.12
    0.6%
    Correlated Neurons
    Index
    P. Corr.
    Cos Sim.
    60
    +0.15
    0.12
    362
    +0.13
    0.13
    608
    +0.12
    0.13
    Negative Logits
    sburg
    -0.73
    fair
    -0.71
    NN
    -0.67
    sov
    -0.65
    hops
    -0.65
     obliged
    -0.64
    former
    -0.63
    pad
    -0.63
    deen
    -0.62
    sat
    -0.62
    POSITIVE LOGITS
     regard
    1.65
     regards
    1.64
    stood
    1.55
    draw
    1.49
     impunity
    1.44
    drawn
    1.39
     respect
    1.31
    standing
    1.21
    holding
    1.11
     utmost
    0.98
    Activations Density 2.621%

    No Known Activations