© 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 EvalsExportsinterp-engineNEWAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-4-31B
    3. 30-RES-MATRYOSHKA-131K
    4. 128339
    Prev
    Next
    INDEX
    Explanations

    * When 'Dou' appears, it's followed by 'bles' (likely part of 'Doubles').3. **TOP_POSITIVE_LOGITS**: * Comissão * ArrivalTime * fläche * permissions * ძ * Posters * schnell * ダーク * zinha * 技巧 These are words/tokens with high positive association with this neuron's activation. They seem very diverse and are from different languages (Portuguese, German, Japanese, etc.). This list might indicate the neuron is activated by a wide range of distinct concepts, or perhaps it's picking up on specific multi-lingual or abstract patterns.4. **TOP_ACTIVATING_TEXTS**: * "English Premier League. The discussions will also focus on whether the format of the domestic Twenty20 group matches can be expanded from two-and-a-half weeks to five weeks" * "County Limerick, Ireland. The club was founded in 1952 and is almost exclusively concerned with hurling with footballers catered for by neighbouring Galbally. Location The club is located in the parish of Garryspillane containing" * "Most cochlear implant (CI) users describe music as a noise-like and unpleasant sound." * "hopeful of leading England in the third ODI.Getty Images England host Sri Lanka at Old Trafford in Manchester for the third One-Day International (ODI) on Wednesday." * "gren The Moon, The Stars And You - CD The long awaited album after “Sentimental Journey”: as a singer and as a trombone player Nils Landgren shows his romantic side." * "The Agency was formed to revitalize this area by eliminating blight and providing affordable housing." * "Unless we use an integrated approach" (context is safe driving) * "Barry Phillips-Moore 6–4, 6–2, 1–6, 6–3 Doubles Tito Vázquez / Guillermo Vil" * "Simple Triage and Rapid Treatment (START) methods." (context is emergency department triage)Now, let's synthesize the findings:* `MAX_ACTIVATING_TOKENS` shows a mix of common words ('the', 'and'), specific entities ('footballers', 'music', 'Agency', 'singer'), and grammatical elements ('Dou').* `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` suggests context-following: 'footballers catered', 'music as', 'singer and', 'Agency was', 'Doubles', 'Rapid Treatment'.* `TOP_POSITIVE_LOGITS` are highly diverse, suggesting the neuron might not be about a single semantic concept but rather a structural or abstract feature captured across languages.* `TOP_ACTIVATING_TEXTS` reveal themes like sports (football, hurling, cricket, tennis), music, and organizational entities/plans (Agency).The pattern seems to be activating around specific nouns, especially those that might be followed by verbs or prepositions that define their role, or referring to specific types of entities or events. The presence of multi-lingual tokens in `TOP_POSITIVE_LOGITS` is a strong hint that the neuron is not purely semantic but might be sensitive to patterns across different linguistic structures or common abstract concepts that manifest in various languages.Looking at `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`:- `footballers catered`- `music as`- `singer and`- `driving`- `Doubles`- `Rapid` (likely part of a phrase like Rapid Treatment)The `TOP_POSITIVE_LOGITS` are key here. Many are from different languages and don't have obvious semantic links across the top activating texts. This suggests the neuron might be capturing a more abstract relationship, perhaps related to noun phrases, role assignment, or specific grammatical structures common across its diverse activations.Consider the sequence `MAX_ACTIVATING_TOKENS` -> `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`:- footballers -> catered- music -> as- singer -> and- Agency -> was- driving -> (context sensitive)- Dou -> bles (doubles)- One -> wasThis shows a pattern of a noun or a specific token being followed by a verb or a word that defines its category/role.The presence of `Comissão`, `ArrivalTime`, `fläche`, `permissions`, `Posters`, `schnell`, `ダーク`, `zinha`, `技巧` suggests that the neuron is not tied to a single English concept. These are very diverse.entity followed by classifier

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    decoderesearch/gemma-4-saes/gemma-4-31b
    Prompts (Dashboard)
    16,384 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     Closer
    -0.06
     ресурсов
    -0.06
     具体
    -0.06
    His
    -0.06
     கட்டுமான
    -0.06
     szolgált
    -0.06
     পরিষে
    -0.06
    binicons
    -0.06
    closer
    -0.06
    唔
    -0.06
    POSITIVE LOGITS
     Comissão
    0.06
    ArrivalTime
    0.06
    fläche
    0.06
    permissions
    0.06
    ძ
    0.06
     Posters
    0.06
     schnell
    0.05
     ダーク
    0.05
    zinha
    0.05
    技巧
    0.05
    Activations Density 0.076%

    No Known Activations