© 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. 1-GEMMASCOPE-2-TRANSCODER-262K
    4. 105278
    Prev
    Next
    INDEX
    Explanations

    % symbols"}, {"explanation": "file paths"}, {"explanation": "code fragments"}, {"explanation": "URL parameters"}, {"explanation": "filename components"}]The `MAX_ACTIVATING_TOKENS` like 'C', 'D', 'Caps', 'd', 'Fg' appear in various contexts within the `TOP_ACTIVATING_TEXTS`.Let's look for common patterns in the `TOP_ACTIVATING_TEXTS` where these tokens might appear prominently:1. `playst%2Caps%2C465&sr=8-1`: This looks like a URL query string. 'Caps' could relate to 'caps' in search terms (like PlayStations) or just be a token. The '%' is important.2. `%CF%81%CE%B7%CF%82_%CE%9D%CF%8C%CF%83%CF%83%CE%B7%CF%82`: This is URL-encoded text. The '%' is clearly a dominant feature.3. `C:%5CUsers%5C...%5CProgramming%5CPetProjects%5CBUSINESS_PROJECTS%5Cbrazilian_ecommerce%5`: This is a file path, heavily using URL encoding (`%5C` for `\`). 'C' and 'Programming' are present.4. `%04d.png`: This is a filename pattern, often used in sequence generation. '%' is present.5. `%03d.png`: Similar to the above.6. `%20Chat%20Files%2FMeet%20%2D%20upc%2Dmsak%2Dpxi%20%E2%80%94%20Mozilla%`: Another URL-encoded string, including `%20` for space. 'ms' is present.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list:- `%`- `%`- `0`- `k`- `Programming`- `need`- `%`- `.`- `.`- `ms`This list strongly reinforces the presence of '%' and numbers, and some specific words/sequences like 'Programming' and 'ms'.The `TOP_POSITIVE_LOGITS` list:- `-`- `:`- `)$`- `2`- `en`- `this`- `l`- `Р`- `ized`This list is a bit more abstract. However, symbols like ':' and '-' are common in file paths and URLs. '2' and 'ized' could be part of sequences.**Synthesizing the pattern:**The most prominent recurring element across `MAX_ACTIVATING_TOKENS`, `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`, and `TOP_ACTIVATING_TEXTS` is the '%' character, often followed by numbers or letters, indicating URL encoding or file path components. The `MAX_ACTIVATING_TOKENS` also include 'C', 'D', 'd', 'Fg' which can appear as filenames or parts of paths. The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list contains specific words like "Programming" and abbreviations like "ms" which are also seen in the `TOP_ACTIVATING_TEXTS` in contexts like file paths.The neuron seems to activate when it sees sequences that are either URL-encoded, parts of file paths, or specific programming-related strings often found in technical contexts. The '%' symbol is a strong indicator.Let's try to capture this:- "URL encoding and file paths" (5 words) - Good, captures two main contexts.- "percent encoding and file paths" (5 words) - More specific about '%'.- "file paths and URL encoded strings" (6 words) - Similar.- "web url encoding and file paths" (6 words) - Emphasizes web context for URL encoding.Given the output includes things like `C:%5CUsers%5C...` which is clearly a file path, and `%CF%81%CE%B7%CF%82...` which is URL encoding, and `ffmpeg -i input.mp4 frame%04d.png` which uses both pathlike structure and encoding (`%04d`), "percent encoding and file paths" seems very fitting. The 'C', 'D', 'd', 'Fg' could be drive letters or parts of filenames/directory names.Let's consider the specific examples:- `playst%2Caps%2C465&sr=8-1`: URL parameters with % encoding.- `C:%5CUsers%5C...%5CProgramming%5C...`: File path with % encoding for slashes.- `frame%04d.png`: Filename pattern using % encoding for sequential digits.The core pattern is the use of '%' to encode characters or represent placeholders in file/path structures and URLs.Possible phrases:- percent encoding symbols- url encoding and paths- file paths and percent encoding- percent encoded characters and file pathsThe prompt asks for what the neuron *detects or predicts*. It's detecting these specific *syntactic structures*."percent encoding and file paths" covers the main aspects well and is concise.Let's check the constraints: 3 to 20 words. "percent encoding and file paths" is 5 words.No capitalization unless proper noun. No "tokens", "patterns", "detects", "predicts". Just the phrase.Let's review the `MAX_ACTIVATING_TOKENS` again: Caps, C, a, D, C, d, Fg, d, d, D.If the token is 'C' and it's followed by `%:\Users\%...`, then 'C' precedes a file path structure.If the token is 'Fg' and it's from `frame%04d.png`, it's part of a filename.The `TOP_POSITIVE_LOGITS` like `:`, `-`, `2`, `ized` also fit into these technical contexts. `2` could be part of a version number or a hex code. `ized` could be a suffix in a word found in a file name or description.The pattern is very strong around '%', file paths, and URL structures.Final check:- Concise (3-20 words): Yes, 5 words.- Single word, phrase, or pattern: Yes, a phrase.- Tokens following/preceding: Covered by the context.- Words starting with sequences: Not directly applicable here, more about structure.- Avoid listing tokens: Yes.- No "words related to", "concepts related to", etc.: Yes.- No "This neuron detects/predicts": Yes.- No "tokens" or "patterns": Yes.- No capitalization unless proper noun: Yes, "percent" and

    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_1_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
     as
    1.16
    的
    1.14
     are
    1.05
    ные
    0.99
    та
    0.98
    ALY
    0.98
    ға
    0.94
    زمانہ
    0.94
    LIM
    0.93
    站点
    0.93
    POSITIVE LOGITS
    -
    1.25
    :
    1.16
    )$
    1.10
    2
    1.08
            
    1.01
    en
    1.01
    this
    1.00
    l
    0.97
    Р
    0.96
    ized
    0.96
    Activations Density 0.001%

    No Known Activations