© 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
    2. Qwen3.5-4B
    3. 15-RES-MATRYOSHKA-65K
    4. 18853
    Prev
    Next
    INDEX
    Explanations

    The `MAX_ACTIVATING_TOKENS` list contains words like "man", "Raider", "-Men", "Wolf".The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list contains words like "Series", "fans", "Album".The `TOP_ACTIVATING_TEXTS` show examples like "Hitman game is a video game series", "Executioner is a monthly men's action-adventure series", "SWAT Kats: The Radical Squadron is an animated series", "Tomb Raider fans", "Bon Iver Album", "X-Men", "The Black Wolf Series".The pattern seems to be related to named entities, particularly fictional characters or series titles, often involving professions or groups, followed by words indicating media type or audience. For example:- "man" (Hitman) -> "game is a video game series"- "er" (Executioner) -> "monthly men's action-adventure series"- "Raider" (Tomb Raider) -> "fans"- "-Men" (X-Men) -> This is a series, though not explicitly followed by a prediction token in this example, it's a strong indicator.- "Wolf" (Black Wolf) -> "Series"The `TOP_POSITIVE_LOGITS` are in non-Latin scripts, which are harder to interpret directly but might indicate a broader concept or a way the model represents certain entities. However, the English texts are clearly showing a pattern of fictional character/group names and media types.Let's look for a concise pattern:- **Characters/Groups:** "man", "-Men", "Wolf"- **Media/Series:** "game is a video game series", "action-adventure series", "animated series", "Series", "Album"- **Audience:** "fans"The neuron seems to highlight titles related to media franchises, especially character names or group names that are part of a series or franchise.Possible explanations:- Titles of fictional series and characters- Names of media franchises and characters- Fictional series and character names- Superhero and gaming franchises- Names of book series and characters- Action/adventure series characters"Fictional series and character names" seems a good fit. It covers "Hitman game", "Executioner", "SWAT Kats", "Tomb Raider", "X-Men", "The Black Wolf Series". "Raider" and "man" are specific parts of these names. The tokens after often confirm it's a series or relates to fans/albums.Fictional series and character names

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    decoderesearch/qwen-3.5-saes/qwen-3.5-4b
    Prompts (Dashboard)
    16,384 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    youtu
    -0.06
    park
    -0.06
    onn
    -0.06
    çļĦ建设
    -0.06
    æĹłåīį
    -0.06
    plane
    -0.06
    æľ¬æ³ķ
    -0.06
    èıĵ
    -0.05
     suit
    -0.05
    fut
    -0.05
    POSITIVE LOGITS
    åħ¥å¾®
    0.06
    immer
    0.05
     nightly
    0.05
    alle
    0.05
    ä½ĵåij³
    0.05
    à¹ģส
    0.05
    è¯Ĺ人
    0.05
    etni
    0.05
    å¾®
    0.05
    ç«ĭåł´
    0.05
    Activations Density 0.002%

    No Known Activations