© 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. Under Peer Review · Sparse Autoencoders for Pythia-70M-Deduped
    5. Pythia-70M-Deduped
    7. Residuals
    9. 4-RES-SM
    11. 13324
    Prev
    Next
    INDEX
    Explanations

    instances of the word "which" and similar conjunctions

    oai_token-act-pair · gpt-4o-miniTriggered by @bot
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    ctigges/pythia-70m-deduped__res-sm_processed/4-res-sm
    Prompts (Dashboard)
    32,768 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    Features
    32,768
    Data Type
    torch.float32
    Hook Name
    blocks.4.hook_resid_post
    Hook Layer
    4
    Architecture
    standard
    Context Size
    128
    Dataset
    EleutherAI/the_pile_deduplicated
    Activation Function
    relu
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Neuron Alignment
    Index
    Value
    % of L₁
    25
    +0.18
    1.0%
    204
    +0.14
    0.8%
    388
    +0.12
    0.7%
    Correlated Neurons
    Index
    P. Corr.
    Cos Sim.
    204
    +0.18
    0.07
    171
    +0.14
    0.06
    388
    +0.12
    0.04
    Negative Logits
    Docket
    -1.71
    ounce
    -1.65
     and
    -1.65
     whom
    -1.63
     watching
    -1.63
     versus
    -1.54
     listening
    -1.47
    -,
    -1.45
    ~~~
    -1.45
     who
    -1.43
    POSITIVE LOGITS
     includes
    2.33
     means
    2.19
     requires
    2.11
     contains
    2.10
     involves
    2.09
     translates
    2.05
     consists
    2.00
     corresponds
    2.00
     happens
    1.92
     leads
    1.86
    Activations Density 0.246%

    No Known Activations