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

    Neuronpedia

    APICircuit TracerNEWSteerSAE EvalsExportsSlackBlogPrivacy & TermsContact
    © Neuronpedia 2025
    Privacy & TermsBlogGitHubSlackTwitterContact
    1. Home
    2. Under Peer Review · Attention SAE Research Paper
    3. GPT2-Small
    4. Attention Out
    5. 1-ATT-KK
    6. 6436
    Prev
    Next
    INDEX
    Explanations

    This attention head attends to "<|endoftext|>" tokens from tokens later in the text, particularly looking at content markers, punctuation or metadata elements.

    oai_attention-head · claude-3-7-sonnet-20250219Triggered by @trainjumperx

    attends to the <|endoftext|> token at the beginning of sequences from subsequent tokens in the sequence.

    oai_attention-head · claude-3-5-sonnet-20240620Triggered by @trainjumperx

    legal and governmental document terminology related to surveillance and data collection

    oai_token-act-pair · gpt-3.5-turboTriggered by @bot
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GPT2-SMALL @ 1-att-kk
    Configuration
    ckkissane/attn-saes-gpt2-small-all-layers/gpt2-small_L1_Hcat_z_lr1.20e-03_l18.00e-01_ds24576_bs4096_dc1.00e-06_rsanthropic_rie25000_nr4_v5.pt
    Prompts (Dashboard)
    36,864 prompts, 128 tokens each
    Dataset (Dashboard)
    Skylion007/openwebtext
    Features
    24,576
    Data Type
    float32
    Hook Name
    blocks.1.attn.hook_z
    Hook Layer
    1
    Architecture
    standard
    Context Size
    128
    Dataset
    Skylion007/openwebtext
    Activation Function
    relu
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Head Attr Weights
    0:0.06
    1:0.05
    2:0.04
    3:0.02
    4:0.05
    5:0.12
    6:0.26
    7:0.12
    8:0.04
    9:0.05
    10:0.04
    11:0.09
    Negative Logits
    YR
    -1.74
     神
    -1.73
     Soda
    -1.72
    Song
    -1.70
     toget
    -1.62
    nir
    -1.61
    ‎
    -1.61
     organise
    -1.60
     Toast
    -1.59
    song
    -1.59
    POSITIVE LOGITS
     FISA
    2.10
     wiret
    2.01
    nsics
    1.97
     databases
    1.95
     spying
    1.86
    missible
    1.79
     affidav
    1.77
     exoner
    1.75
     subpoen
    1.74
     confidential
    1.72
    Activations Density 0.010%

    No Known Activations