© 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. Gemma-3-27B-IT
    3. 15-GEMMASCOPE-2-TRANSCODER-262K
    4. 15686
    Prev
    Next
    INDEX
    Explanations

    The MAX_ACTIVATING_TOKENS list only contains `+` characters, which aren't informative on their own.Let's look at TOKENS_AFTER_MAX_ACTIVATING_TOKEN:`;`, `Shift`, `R`, `Shift`, `J`, `Shift`, `Shift`, `Shift`The presence of 'Shift' repeatedly, along with other keys like ';' and 'R' and 'J', strongly indicates keyboard shortcuts.Let's look at TOP_ACTIVATING_TEXTS:- `Ctrl+Alt+Shift+S`- `Cmd+;`- `Ctrl + Shift + Delete`- `Ctrl+R`- `Ctrl + H`- `Cmd + Shift + H`- `Ctrl+Shift+V`- `Cmd+Shift+V`- `Ctrl+Shift+X`- `Cmd+Shift+X`- `Ctrl+J`- `Cmd+J`- `Ctrl+Shift+B`- `Cmd+Shift+B`- `Ctrl+Shift+N`- `Cmd+Shift+O`This list overwhelmingly shows examples of keyboard shortcuts, often involving modifier keys like Ctrl, Cmd, Alt, and Shift, followed by another key (like S, ;, Delete, R, H, V, X, B, N, O, J).Let's look at TOP_POSITIVE_LOGITS:`gacche`, `ရော`, `vutta`, `GONDOR`, `strept`, `捜`, `abilă`, `ının`, `manuss`, `\{-\`These seem like random tokens, but they are the most likely words/tokens associated with this neuron. They don't immediately form an obvious pattern related to the keyboard shortcuts. They might be related to the *context* where these shortcuts are mentioned (e.g., specific software names or commands, foreign languages), but the dominant pattern is the keyboard shortcuts.The prompt asks me to find a pattern. The strongest, most consistent pattern across TOP_ACTIVATING_TEXTS and suggested by TOKENS_AFTER_MAX_ACTIVATING_TOKEN is keyboard shortcuts. The MAX_ACTIVATING_TOKENS are uninformative here, but the tokens *after* them reinforce the keyboard shortcut idea. The TOP_POSITIVE_LOGITS might be what the neuron *associates* with these shortcuts, perhaps specific applications or contexts.The goal is a concise explanation (3-20 words) of what the neuron detects or predicts by finding patterns.The neuron clearly relates to keyboard shortcuts.Let's reconsider what "MAX_ACTIVATING_TOKENS" and "TOKENS_AFTER_MAX_ACTIVATING_TOKEN" mean in the context of a

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-it/transcoder_all/layer_15_width_262k_l0_small_affine
    Prompts (Dashboard)
    238,145 prompts, 512 tokens each
    Dataset (Dashboard)
    lmsys + oasst1
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    آ
    0.32
    р
    0.30
    ch
    0.30
    re
    0.30
    ن
    0.30
    وی
    0.29
    له
    0.28
    ksi
    0.28
    پ
    0.28
    на
    0.27
    POSITIVE LOGITS
     gacche
    0.33
    ရော
    0.29
     vutta
    0.28
    GONDOR
    0.28
     strept
    0.26
    捜
    0.26
    abilă
    0.25
    ının
    0.25
     manuss
    0.25
    \{-\
    0.25
    Activations Density 0.008%

    No Known Activations