© 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 EvalsExportsinterp-engineNEWAPI Community BlogPrivacy & TermsContact
    1. Home
    3. Under Peer Review · Attention SAE Research Paper
    5. GPT2-Small
    7. Attention Out
    9. 3-ATT-KK
    11. 13252
    Prev
    Next
    INDEX
    Explanations

    events or incidents related to controversial or significant moments

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

    events related to significant societal issues and controversies

    oai_token-act-pair · gpt-4o-miniTriggered by @bot
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GPT2-SMALL @ 3-att-kk
    Configuration
    ckkissane/attn-saes-gpt2-small-all-layers/gpt2-small_L3_Hcat_z_lr1.20e-03_l19.00e-01_ds24576_bs4096_dc1.00e-06_rsanthropic_rie25000_nr4_v9.pt
    Prompts (Dashboard)
    36,864 prompts, 128 tokens each
    Dataset (Dashboard)
    Skylion007/openwebtext
    Features
    24,576
    Data Type
    float32
    Hook Name
    blocks.3.attn.hook_z
    Hook Layer
    3
    Architecture
    standard
    Context Size
    128
    Dataset
    Skylion007/openwebtext
    Activation Function
    relu
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Head Attr Weights
    0:0.02
    1:0.07
    2:0.09
    3:0.05
    4:0.01
    5:0.07
    6:0.05
    7:0.08
    8:0.22
    9:0.07
    10:0.12
    11:0.11
    Negative Logits
     instead
    -1.45
     nonetheless
    -1.35
     accordingly
    -1.18
     ensu
    -1.09
    endif
    -1.08
     squid
    -1.07
     Vader
    -1.07
     contrad
    -1.07
     whilst
    -1.06
     backwards
    -1.06
    POSITIVE LOGITS
    idelines
    1.16
    ittal
    1.15
    ynt
    1.13
    assault
    1.11
    ynthesis
    1.07
    ounter
    1.04
    anity
    1.04
    artz
    1.03
     scandal
    1.03
    abuse
    1.02
    Activations Density 0.070%

    No Known Activations