© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    Natural Language
    Autoencoders
    NEW
    Assistant AxisNEWCircuit TracerUPDATESteerSAE EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-12B
    3. 24-GEMMASCOPE-2-RES-16K
    4. 4908
    Prev
    Next
    INDEX
    Explanations

    The `MAX_ACTIVATING_TOKENS` list contains tokens like `/`, `##`, `一`, and `'` which often appear at the start of lines, headings, or as separators.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list shows tokens such as `m`, `Modules`, `能源`, `checkpoint`, and `current` following these initial tokens.The `TOP_POSITIVE_LOGITS` list has terms like `python`, `一` (one), `中国` (China), `反` (against/anti), `中` (middle/in), `对` (to/for/correct).The `TOP_ACTIVATING_TEXTS` show examples of code, markdown, and technical descriptions, sometimes in Chinese and English.Looking for a pattern:- `一` is common in both `MAX_ACTIVATING_TOKENS` and `TOP_POSITIVE_LOGITS`.- `##` and `/` often start blocks of text or code.- `Modules`, `能源` (energy), `checkpoint`, `current`, and `python` suggest technical or informational content.- The Chinese characters `一`, `中`, `对`, `反` suggest the neuron might be involved in processing Chinese text, possibly related to classification or identification within that text.- The presence of `python` and technical terms like `checkpoint` points towards code or technical documentation.Let's synthesize these.The neuron seems to activate on technical content, sometimes in Chinese. The initial tokens `/` and `##` might indicate the start of certain types of data blocks. The word `一` (one) is frequently seen. `Modules` and `checkpoint` are technical terms. `能源` (energy) is another technical term.The phrase should be concise and reflect a pattern."Modules 能源 checkpoint" is too literal."一" is very basic.Consider the examples:- `## 编程和 Prolog 的联系和区别`: `##` is a markdown heading. `一` is a common Chinese number. `编程` (programming) is technical.- `Modules[#](#modules "Permalink to this headline")`: `Modules` is the word, `##` is markdown.- `能源消耗量`: `能源` (energy) is a technical term.- `gradient_checkpointing`: technical term.It looks like it's identifying technical terms, possibly within structured text like markdown or code comments. The chinese terms might indicate it's sensitive to specific Chinese technical domains or general categorization of Chinese text."chinese technical terms" - too generic."technical content in chinese" - better."technical chinese words" - also good.Let's look at `MAX_ACTIVATING_TOKENS` again: `/`, `##`, `一`, `'`, `_`, `2`, `the`.`##` and `/` are structural. `一` is a number/separator. `'` and `_` are common in code. `2` is a number. `the` is common English.The `TOP_POSITIVE_LOGITS` having `python`, `中国`.technical words and chinese

    np_acts-logits-general · gemini-2.5-flash-lite

    The neuron is primarily detecting numeric tokens (i.e. number literals) in the text.

    oai_token-act-pair · o4-miniTriggered by @jyhe0408
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-12b-pt/resid_post/layer_24_width_16k_l0_medium
    Prompts (Dashboard)
    392,802 prompts, 256 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     También
    1.23
     Puedes
    1.23
     Tamaño
    1.20
     ተጨማሪ
    1.17
     GALLERY
    1.16
     tambien
    1.15
     Muchos
    1.14
     Muitos
    1.14
     Detalles
    1.13
     Alguns
    1.13
    POSITIVE LOGITS
    <0xE3>
    1.16
    :
    1.12
    中
    1.06
    即
    1.03
    ;
    1.02
    自
    0.95
    对
    0.95
    后
    0.95
    !
    0.95
     zhong
    0.94
    Activations Density 0.000%

    No Known Activations