© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it
    Neuronpedia
    Jacobian LensNEW
    Natural Language
    Autoencoders
    MAY 26
    Assistant AxisJAN 26Circuit TracerAPR 26SteerSAE EvalsExportsinterp-engineNEW Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-27B-IT
    3. 15-GEMMASCOPE-2-TRANSCODER-262K
    4. 13224
    Prev
    Next
    INDEX
    Explanations

    The neuron seems to detect words related to identity or specific entities, especially when followed by specific characters or continuations: "Nostromo", "Barbarian", "OC", or possibly a reference to the NAACP.Let's re-evaluate with the examples:**MAX_ACTIVATING_TOKENS**: ost, Barb, arian, OC, ac, image, ac, ac, ac, ac**TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: romo, arian, *, -, p, ")", p, p, p, p**TOP_POSITIVE_LOGITS**: ké, nsec, 𝚢, 小姐, lsch,έλ, zal, ρό, finalists, ئی**TOP_ACTIVATING_TEXTS**:* "Nostromo" mission. (ost -> romo)* "Conan the Barbarian!" (Barb -> arian)* "Swiss Coffee OC-45" (OC -> *) (OC is followed by a '*' in the text, which is in TOKENS_AFTER_MAX_ACTIVATING_TOKEN)* References to NAACP (implied by the context after "ac", which might be a shortened form or a typo of something else). The repetition of "ac" and "p" might indicate a structured list or formatting.The commonality seems to be specific nouns or classifications, followed by specific continuations (romo, arian, *, or perhaps a list structure after 'ac')."Nostromo", "Barbarian", "OC-45". These are proper nouns or specific classifications.The "ac" repeated with "p" after it might indicate list items or specific formatting.Let's focus on the most prominent ones:- `ost` + `romo` -> `Nostromo`- `Barb` + `arian` -> `Barbarian`- `OC` + `*` -> `OC` (specifically contextually followed by formatting, "OC-45")The TOP_POSITIVE_LOGITS are very diverse, ranging from Chinese characters '小姐' to Greek 'έλ' and 'ρό', and 'finalists'. This is confusing.Let's re-examine the TOP_ACTIVATING_TEXTS for the 'ac' tokens.The 'ac' tokens appear in:1. "Swiss Coffee OC-45:" (Here, 'OC' is the max activating token, '45' follows, but '*' is in the TOKENS_AFTER_MAX_ACTIVATING_TOKEN list. The prompt says tokens *after* max activating token. So `OC` -> `*` is the match).2. "learning more about why this word is harmful: * **The NAACP:** [https://naacp.org/find-resources/answers-to-common-questions/understanding-the-n-word](" - this contains "Understanding the N-word", NAACP. The text is cut off.3. "learning more about why that word is offensive, here are some resources: * **The NAACP:** [https://naacp.org/find-resources/answers-to-frequently-asked-questions-about-the-n"4. "showingupforracialjustice.org/) * **The NAACP:** [https://naacp.or"5. "showingupforracialjustice.org/) * **The NAACP:** [https://naacp.org/"The letters `ac` appear as part of `NAACP`. The `ac` in `NAACP` is activating. The tokens following `ac` in `NAACP` aren't explicitly listed.However, we see `ac` repeated in MAX_ACTIVATING_TOKENS and 'p' repeatedly in TOKENS_AFTER_MAX_ACTIVATING_TOKEN.This suggests a pattern related to `ac` followed by `p` (or list items).Let's try to combine the observable patterns:- `ost` -> `romo` (Nostromo)- `Barb` -> `arian` (Barbarian)- `OC` -> `*` / `45` (OC-45)- `ac` -> `p` (potentially related to NAACP or lists)The prompt asks for a *concise* explanation (3-20 words).The most striking patterns are "Nostromo" and "Barbarian". The "OC" is also clear. The "ac" with subsequent 'p' is less clear but present.Possible

    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
    Negative Logits
    ములో
    0.11
     sinister
    0.11
    いわ
    0.11
     apud
    0.11
    得分
    0.10
     करतो
    0.10
    Quanto
    0.10
     asymmetrical
    0.10
     descrito
    0.10
    ворю
    0.10
    POSITIVE LOGITS
    ké
    0.14
    nsec
    0.13
    𝚢
    0.13
    小姐
    0.12
    lsch
    0.12
    έλ
    0.11
    zal
    0.11
    ρό
    0.11
     finalists
    0.11
    ئی
    0.11
    Activations Density 0.007%

    No Known Activations