© 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. Apollo Research · Taylor · Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning
    5. GPT2-Small
    7. Residual Similar CE Local
    9. 6-RES_SCL-AJT
    11. 4181
    Prev
    Next
    INDEX
    Explanations

    harsh or critical language and statements in the text

    oai_token-act-pair · gpt-3.5-turbo
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    neuronpedia/gpt2-small__res_scl-ajt/6-res_scl-ajt
    Prompts (Dashboard)
    12,288 prompts, 128 tokens each
    Dataset (Dashboard)
    Skylion007/openwebtext
    Features
    46,080
    Data Type
    torch.float32
    Hook Point
    blocks.6.hook_resid_pre
    Architecture
    standard
    Context Size
    128
    Dataset
    apollo-research/Skylion007-openwebtext-tokenizer-gpt2
    Hook Point Layer
    6
    Activation Function
    relu
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Neuron Alignment
    Index
    Value
    % of L₁
    289
    +0.11
    0.6%
    311
    +0.10
    0.5%
    183
    +0.09
    0.5%
    Correlated Neurons
    Index
    P. Corr.
    Cos Sim.
    418
    +0.11
    0.11
    289
    +0.10
    -0.03
    35
    +0.09
    0.06
    Negative Logits
    ually
    -0.81
    duct
    -0.74
    iven
    -0.73
    elta
    -0.72
    ovember
    -0.71
    phis
    -0.70
    xxxxxxxx
    -0.69
    plex
    -0.68
    agara
    -0.67
    eer
    -0.67
    POSITIVE LOGITS
     harshly
    0.88
     harsh
    0.87
     hars
    0.86
     harsher
    0.83
     reception
    0.77
     criticism
    0.74
     parting
    0.73
     opinions
    0.71
     criticisms
    0.71
     rhetoric
    0.70
    Activations Density 10.351%

    No Known Activations