© 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
    3. 31-GEMMASCOPE-2-RES-262K
    4. 229692
    Prev
    Next
    INDEX
    Explanations

    Controls2. Model Binding Part 1 - Selecting Data3. Model Binding Part 2 - Filtering4. rhetorically found, our download strategies in regenerative medicine integrating will see accepted and the tag will continue reviewed. know You for Helping us" - `Controls2.` followed by `Model`. Controls2. Model Binding Part 1 - Selecting Data3. Model Binding Part 2 - Filtering4. rhetorically found, our download strategies in regenerative medicine integrating will see accepted and the tag will continue reviewed. know You for Helping us"Looking at MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN:- `2` is followed by `0`, `found`, `2`.- `.` is followed by `0`, `Museum`, `info`, `found`, `open`.- `,` is followed by `Museum`, `info`.Looking at TOP_POSITIVE_LOGITS:- `Controls2.` is followed by `Model`.- `Templates`- `Template`- `diagram`- `TimeDict`The pattern `Controls` followed by a number (`2`) and then `.` is seen in one of the TOP_ACTIVATING_TEXTS.The `MAX_ACTIVATING_TOKENS` list contains `2`, `.`, and `,`.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list contains `Museum`, `info`, `found`, `open`, `0`.The phrase "Controls" followed by a number and a period seems to be a strong indicator. Then, what follows is often a specific category or item.The examples show numbers, especially followed by periods or commas, which often precede certain words like "Museum", "info", "found", "open".The TOP_POSITIVE_LOGITS suggest elements related to structured data or lists (Templates, diagram,TimeDict).The pattern seems to be numeric identifiers followed by categories or descriptive text.Consider "Controls2. Model Binding Part 1 - Selecting Data3. Model Binding Part 2 - Filtering4."The neuron is looking for numbered list items.The `2` and `.` in MAX_ACTIVATING_TOKENS are very prominent.The TOKENS_AFTER shows items like `Museum`, `info`, `found`, `open` after these numbers/periods.Let's try to capture the essence: numbers followed by periods or commas leading to specific content.The positive logits also suggest structured items like `Templates` and `diagram`.How about focusing on the numbers and what follows them?`2` followed by `.` implies a list item.`.` followed by `info` or `Museum` or `found`.The `Controls2.` example is strong. It implies a pattern of "Identifier" + Number + Punctuation.`Controls2.` is followed by `Model`.`2.` could be followed by `Templates` or `diagram`.The idea of numbered lists seems to be present.Let's consider the simplest, most recurring element: numbers followed by punctuation. `2` and `.` are very frequent.What follows these? `0`, `found`, `open`, `info`, `Museum`.The structure `Number + .` or `Number + ,` is common.The `TOP_POSITIVE_LOGITS` are `Templates`, `Template`, `diagram`. These could be items found *after* such identifiers.Let's re-examine `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `Museum`, `info`, `found`, `open`.These are nouns or verbs indicating information or existence.Combining `2`, `.`, `,` with `Templates`, `diagram`, `info`, `Museum`.It looks like specific items or categories, often following a numerical or structural marker.The neuron seems to be identifying specific entries or details, often indicated by enumeration or punctuation.The phrase "Number followed by specific detail" is too long."numbered entries and descriptions" is still too long, more than 20 words.Let's look at `MAX_ACTIVATING_TOKENS`: `2`, `.`, `,`, `ically`.`TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `0`, `Museum`, `info`, `2`, `found`, `0`, `open`.The `2` and `.` are key. What follows them? `0`, `found`, `open`, `info`.The `TOP_POSITIVE_LOGITS` are `Templates`, `diagram`.The most specific commonality is numeric indicators and then labels or types.Could it be specific cataloging labels?Let's focus on the `Controls2.` example. This is `Identifier` + `Number` + `.`The `Max_activating_tokens` `2` and `.` are strongly present.The `Tokens_after` `Museum`, `info`, `found`, `open` are concrete nouns/states.The `Top_positive_logits` `Templates`, `diagram` are also concrete nouns.The neuron seems to detect specific items or categories that appear after numerical enumeration or structural markers."numbered items followed by names" - 5 words."entries identified by number and label" - 6 words.Consider the most direct link: `2.` is followed by `found` or `0` which might be part of a larger structure like enumeration.`Templates`, `diagram` are good candidates for what follows such markers.Let's try to find a very specific pattern.`Controls2. Model``2018. IEEE``No. houses``, Aguilar 4``, Monfort 2`The `2` and `.` are prominent.`TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `Museum`, `info`, `found`, `open`.These are common words that appear after lists or enumerations."enumerated items and details" -

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-pt/resid_post/layer_31_width_262k_l0_medium
    Prompts (Dashboard)
    392,802 prompts, 256 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Negative Logits
    uir
    0.50
    ova
    0.49
    out
    0.48
    irc
    0.48
    ave
    0.43
    ie
    0.42
    io
    0.42
     fw
    0.42
    sofar
    0.42
    vre
    0.42
    POSITIVE LOGITS
    TimeDict
    0.47
     ہوا۔
    0.47
     پھیل
    0.46
    聝
    0.46
     Templates
    0.46
    Templates
    0.46
    Template
    0.44
    diagram
    0.44
    崙
    0.44
    彠
    0.44
    Activations Density 0.001%

    No Known Activations