© 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
    2. Llama3.3-70B-IT
    3. 50-RESID-POST-GF
    4. 22897
    Prev
    Next
    INDEX
    Explanations

    assistant\n\nNSFW"- "pussy.<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"Fuck me harder"- "sexual fantasies. You describe the scenario and what I see"- "a mosquito<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nIt's not uncommon"The structure `<|start_header_id|>assistant<|end_header_id|>` followed by content from the assistant seems to be a recurring pattern. The "TOP_POSITIVE_LOGITS" also contain gibberish which is often seen when a neuron is not strongly aligned to specific language but rather to specific structural/control tokens or sequences.The `<|start_header_id|>` token is often followed by `assistant`. The `TOP_POSITIVE_LOGITS` list does not seem to offer clear semantic meaning, but hints at unusual characters or possibly encoded stuff. The `TOP_ACTIVATING_TEXTS` show the `<|start_header_id|>assistant<|end_header_id|>` sequence consistently preceding content from the assistant.Therefore, the rule seems to be related to the start of an assistant's turn.Explanation pattern: `<|start_header_id|>` followed by `assistant`. This suggests generating assistant turns.Let's check the constraints:- 3 to 20 words: "assistant turn start" is 3 words.- Phrase, not sentence: Yes.- Specific: Yes, refers to a structural element.- Avoid forbidden phrases: Yes.- Capitalization: `assistant` should not be capitalized unless it's a proper noun, but here it's a role indicator. "turn start" is lowercase.The pattern seems to specifically target the transition into an assistant's response.Let's refine:The input tokens are `<|start_header_id|>` and `the`.The tokens after them are often `assistant` or `scenario`.The texts frequently contain `<|start_header_id|>assistant<|end_header_id|>` and then assistant content.This neuron seems to activate when the model is about to output text *as* the 'assistant', often after a specific header token.The core pattern is the start of an "assistant" response, often following a special token.Refined explanation: "assistant turn start" or "start of assistant response"."assistant turn start" fits the word count and spirit. assistant turn start

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With LLAMA3.3-70B-IT @ 50-resid-post-gf
    Configuration
    Goodfire/Llama-3.3-70B-Instruct-SAE-l50/Llama-3.3-70B-Instruct-SAE-l50.pt
    Prompts (Dashboard)
    10,000 prompts, 128 tokens each
    Dataset (Dashboard)
    lmsys/lmsys-chat-1m
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    ï½ľ
    -0.10
     Mats
    -0.08
     carte
    -0.08
     Erin
    -0.08
    jar
    -0.08
    insk
    -0.08
    Getty
    -0.08
    ilies
    -0.08
    emo
    -0.08
    Intialized
    -0.08
    POSITIVE LOGITS
     Bilg
    0.10
    ekk
    0.10
     nonatomic
    0.09
    ditor
    0.08
    è¾ij
    0.08
    _counters
    0.08
     ìĸ´ëĬIJ
    0.08
    ÏĥειÏĤ
    0.08
     ÎļαÏĦηγοÏģία
    0.08
    AJOR
    0.08
    Activations Density 0.145%

    No Known Activations