© 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-2-27B
    3. 34-GEMMASCOPE-RES-131K
    4. 69903
    Prev
    Next
    INDEX
    Explanations

    fr / followed by >*Self-reflection:*The user wants a concise explanation (3-20 words) of a neuron's behavior based on provided lists.I need to find a pattern across `MAX_ACTIVATING_TOKENS`, `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`, `TOP_POSITIVE_LOGITS`, and `TOP_ACTIVATING_TEXTS`.Looking at `MAX_ACTIVATING_TOKENS`: `fr`, `https`, `Am`, `in`, `ESI`.Looking at `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `fr` is followed by `>`. `https` is followed by `://`. `ESI` is followed by `,`. `Am` is followed by `.`.Looking at `TOP_ACTIVATING_TEXTS`:- `https://simtk.org/frs/?group_id=1916` -> `fr` followed by `>` (after `simtk.org/`)- `https://dx` -> `https` followed by `://`- `gael.guennebaud@inria.fr>` -> `fr` followed by `>` (e.g., in email addresses)- `$^{*}$ ESI, A-1090 Wien, Boltzmangasse 9, Austria.` -> `ESI` followed by `,`- `j. Opt. Soc. Am. A27, 2169 (2010)]` -> `Am` followed by `.`- `https://leetcode.com/problems/diagonal-traverse/` -> `https` followed by `://`The strongest pattern seems to be `fr` followed by `>` or a separator, often related to URLs or code snippets. `https` followed by `://` is also strong for web links. `ESI` followed by a comma seems specific to an Austrian institute. `Am` followed by a period is common in abbreviations.The `TOP_POSITIVE_LOGITS` like "Matrix", "gram" suggest a mathematical or academic/scientific context, which aligns with the URLs and potential references.Considering the requirement for specificity and pattern identification, "fr / followed by >" captures a recurring and specific sequence seen in the data, particularly relating to abbreviations or parts of URLs. Given the email addresses and URLs, `fr` often precedes a closing bracket `>` or other separators, especially in code comments or technical documentation.Let's re-evaluate:`MAX_ACTIVATING_TOKENS`: `fr`, `https`, `Am`, `in`, `ESI``TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `fr` -> `>`, `https` -> `://`, `Am` -> `.`, `ESI` -> `,``TOP_ACTIVATING_TEXTS` examples:- `simtk.org (<https://simtk.org/frs/?group_id=1916>)` -> `frs` is part of a URL segment, `gr` is followed by `>`- `gael.guennebaud@inria.fr>` -> `fr` followed by `>` (email address)The neuron seems to activate on URL components or email addresses and then predict things that follow them, like separators or link endings. The `fr` token, often part of an email address or a URL path segment, and its common follower `>` or `://` seem to be a strong signal."fr followed by >" is 4 words. It is specific."fr > or ://" covers more but is less specific to the immediate token."fr / or ://" is similar.Let's focus on what `fr` is doing. It appears multiple times in `MAX_ACTIVATING_TOKENS` and in `TOP_ACTIVATING_TEXTS` as part of email addresses (`@inria.fr>`) and URLs (`/frs/`). In these contexts, it's often followed by a closing bracket `>` or other delimiters.What about `Am`? `j. Opt. Soc. Am. A27` -> `Am.`What about `ESI`? `-- ESI, A-1090 Wien...` -> `ESI,`The phrase "fr / followed by >" seems too specific if other patterns exist.Let's try to find a broader pattern.The `TOP_POSITIVE_LOGITS` include "Matrix", "gram", which are technical/mathematical.The `TOP_ACTIVATING_TEXTS` are very technical: scientific papers, code, URLs, email addresses.The neuron seems to be identifying "technical fragments" and what follows them.The most prominent `MAX_ACTIVATING_TOKENS` contributing to a clear sequence are `fr`, `https`, `Am`, `ESI`.- `fr` -> `>` (in email/URL context)- `https` -> `://`- `Am` -> `.` (in abbreviations)- `ESI` -> `,` (in institute context)The repeated `fr` token in `MAX_ACTIVATING_TOKENS` and its common follower `>` in `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` and `TOP_ACTIVATING_TEXTS` is the most

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-27b-pt-res/layer_34/width_131k
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     generated
    -0.82
    eventos
    -0.78
    لمة
    -0.77
    psfrag
    -0.77
     думал
    -0.76
     directly
    -0.74
    textnormal
    -0.74
    Several
    -0.73
     several
    -0.72
     swiftly
    -0.72
    POSITIVE LOGITS
     Matrix
    0.86
    matrix
    0.86
    gram
    0.82
    unk
    0.81
     FEB
    0.77
    ården
    0.77
    луб
    0.77
    Matrix
    0.76
    lift
    0.75
     Playa
    0.74
    Activations Density 0.009%

    No Known Activations