© 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-3-12B-IT
    3. 12-GEMMASCOPE-2-RES-16K
    4. 1950
    Prev
    Next
    INDEX
    Explanations

    your explanation if one cannot be found, but I am confident I can find one.Let's break down the provided lists to understand what this neuron is capturing.1. **MAX_ACTIVATING_TOKENS**: * `T`, `A`, `l`, `r`, `e`, `t`, `B`, `R`, `E`, `ta` * These are primarily single letters, sometimes uppercase, sometimes lowercase. They could be starting letters of words, or possibly parts of acronyms or specific codes.2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: * `.`, `.`, `\'`, `b`, `6`, `Strawberry`, `-`, `.`, `|`, `-` * These follow the activating tokens. We see punctuation (`.`, `'`, `-`, `|`), a lowercase letter (`b`), a number (`6`), and a word (`Strawberry`).3. **TOP_POSITIVE_LOGITS**: * `ﺎ`, `데`, `ᅨ`, `viện`, `inner`, `stricken`, `ﻔ`, `ﻮ`, `лён`, `হস্তে` * This list is very diverse. It contains characters from different scripts (Arabic, Korean, Vietnamese, possibly Cyrillic and Bengali).Okay, let's analyze the provided data to explain the neuron's behavior.**MAX_ACTIVATING_TOKENS**: T, A, l, r, e, t, B, R, E, ta**TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: ., ., ', b, 6, Strawberry, -, ., |, -**TOP_POSITIVE_LOGITS**: ﺎ, 데, ᅨ, viện, inner, stricken, ﻔ, ﻮ, лён, হ**TOP_ACTIVATING_TEXTS**:* "W.I.T.C.H episode "G is for Garbage""* "Chief Information Officer and write 3 S.M.A.R.T. goals on creating an IT Strategic Plan"* "**P**urple (P is the first letter, and Purple starts with P) * **S**trawberry -> **R**ed (S is the first letter, and Red starts with R)"* "Can you spell IABot for me?"* "J.A.R.V.I.S."Observing the `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`, we see single letters appearing.The `TOP_ACTIVATING_TEXTS` provide crucial clues:* "G is for Garbage"* "S.M.A.R.T."* "**P**urple ... starts with P"* "**S**trawberry -> **R**ed"* "IABot"* "J.A.R.V.I.S."The neuron seems to activate when encountering letters, especially those that are either stand-alone (like in initialisms or acronyms) or are presented as the start of a word or phrase, often followed by punctuation or context that defines them. The `TOP_POSITIVE_LOGITS` are diverse, suggesting the neuron is not tied to a specific language but rather the *structure* of how certain linguistic units (like single letters or short sequences) are presented.The pattern is the appearance of single letters, often emphasized, as standalone units or significant starting points within texts.Explanation: **single letters or initialisms**single letters or initialisms

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-12b-it/resid_post/layer_12_width_16k_l0_medium
    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
    其他
    1.46
    و
    1.41
    u
    1.22
    على
    1.20
    ال
    1.18
    auern
    1.17
    قابل
    1.12
    ו
    1.12
     stargazerCount
    1.11
    kval
    1.10
    POSITIVE LOGITS
    ﺎ
    1.22
    데
    1.16
    ᅨ
    1.05
     viện
    1.04
    이너
    1.04
    stricken
    1.04
    ﻔ
    1.03
    ﻮ
    1.02
    лён
    1.02
     হস্তে
    1.01
    Activations Density 0.035%

    No Known Activations