© 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
    2. Andy Arditi · Finding Misaligned Persona Features in Open-Weight Models
    3. Llama3.1-8B-IT (Instruct)
    4. Resid Post - 131k
    5. 11-RESID-POST-AA
    6. 125938
    Prev
    Next
    INDEX
    Explanations

    threat

    np_max-act · gemini-2.0-flash

    The neuron activates on language expressing threats of harm or suicide (e.g. verbs like “amenazar,” “threaten,” “kill”), so it’s looking for threatening or violent verbal cues.

    oai_token-act-pair · o4-miniTriggered by @xinyanhu8
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    andyrdt/saes-llama-3.1-8b-instruct/resid_post_layer_11/trainer_1
    Dataset (Dashboard)
    Various
    Features
    131,072
    Data Type
    float32
    Hook Name
    blocks.11.hook_resid_post
    Architecture
    standard
    Context Size
    1,024
    Dataset
    monology/pile-uncopyrighted
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    wnd
    -0.07
    -Muslim
    -0.07
    伍
    -0.06
     اختی
    -0.06
     Tables
    -0.06
    appId
    -0.06
     oceans
    -0.06
     filesystem
    -0.06
    Browser
    -0.06
     것
    -0.06
    POSITIVE LOGITS
     threatened
    0.08
    "&
    0.07
     mongodb
    0.06
    ��
    0.06
    ,t
    0.06
    ikut
    0.06
     bud
    0.06
     lie
    0.06
     threatening
    0.06
     adolescence
    0.06
    Activations Density 0.008%

    No Known Activations