© 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-3-27B
    3. 31-GEMMASCOPE-2-RES-262K
    4. 83820
    Prev
    Next
    INDEX
    Explanations

    releases and"},"Flush the toilets"},"nursing services or emergency services"}, "Walmart eBooks by partnering with Kobo can boast about pulling a strong library with at"}, "and imaging reports. The treating doctor can explain what the test results mean and whether other tests will be needed in future. Some women find"}, "and the ones who are going to. All of us make right and thought-out decisions that sometimes don"Let's revisit the samples:- MAX_ACTIVATING_TOKENS: `training`, `was`, `essay`, `releases`, `Flush`, `services`, `the`, `customers`, `reports`, `who`- TOKENS_AFTER_MAX_ACTIVATING_TOKEN: `he`, `fine`, `.`, `and`, `the`, `third`, `via`, `.`, `’`Let's map them:1. training -> he2. was -> fine3. essay -> .4. releases -> and (text shows "releases and", but the list has "the" after "releases". This is an inconsistency, let's prioritize the text/logic.)5. Flush -> the (text shows "Flush the")6. services -> the (text shows "nursing services or emergency services", then "The course...")7. the -> third (text shows "on my desk. I don’t know that that" for "the", and "third" appears later in another context)8. customers -> via (text shows "Walmart eBooks by partnering with Kobo can boast about pulling a strong library with at") - "customers via" is not in the text.9. reports -> . (text shows "imaging reports. The treating...")10. who -> ’ (text shows "decisions that sometimes don") - "who" is not in the text in closest proximity.The TOP_POSITIVE_LOGITS are in various languages and refer to concepts like "released", "afterwards", "first", "with", "already", "later", "sentences/propositions".Looking at `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` pairs, some patterns emerge:- `training he`- `was fine`- `releases and` (from text)- `Flush the`- `services or` (from text analysis)- `reports. The`The TOP_POSITIVE_LOGITS suggest a connection to timing (later, afterwards, already) and completion/release (вышла, предложения).The `MAX_ACTIVATING_TOKENS` list itself contains words like `training`, `essay`, `releases`, `services`, `reports` which relate to tasks, documents, or official functions. The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` shows common continuations like `he`, `fine`, `.`, `and`, `the`.The most specific pattern is `was fine`. Also, `Flush the toilets` and `nursing services or emergency services` point towards activities or states.The overall theme seems to be about states, actions, or outcomes, frequently followed by a continuation.`was fine` is a strong indicator.The foreign words in `TOP_POSITIVE_LOGITS` often relate to "released", "first", "later", "already".Let's look for a commonality:- `training he`- `was fine`- `essay.`- `releases and`- `Flush the`- `services or`- `reports.`Many of these are statements of fact or completion. "was fine" is a state. "releases and" indicates an event occurring. "Flush the" is an action. "services or" differentiates options. "reports." indicates an outcome.The foreign words are very suggestive of temporal information or status updates."вышла" (released)"впоследствии" (subsequently)"уже" (already)"později" (later)This neuron might be picking up on statements of completion or subsequent events."was fine" fits."releases and" fits."training he had previously" fits.Let's focus on the `MAX_ACTIVATING_TOKENS` list and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list in conjunction.`was fine` is the most striking phrase.`training he` suggests prior action.`reports.` suggests completion.Consider phrases that capture status updates or events that have occurred or are occurring.If we focus on the direct pairs: "was fine", "releases and", "Flush the", "reports.".The `TOP_POSITIVE_LOGITS` point to verbs or adverbs of time/state like: released, afterwards, already, later.This neuron seems to be about states, events, and their outcomes or timing. "was fine" is a strong clue.Other candidates:- states and outcomes- events and their continuation- completed actions or statesLet's try to refine based on the provided lists again, prioritizing direct pairs or strong overall trends.`was fine` is very specific. `releases and` also. `Flush the` also.The key is that it finds *patterns in lists*. The input lists are tokens.`was` followed by `fine` is a pattern.`releases` followed by `and` (from the text, despite list mismatch).`Flush` followed by `the`.`reports` followed by `.`.This neuron seems to detect common phrase structures indicating states or events."states or events" is too general."detected phrases like 'was fine'" is too meta.Let's look at the positive logits: "released", "afterwards", "already", "later".This hints at temporal aspects or state changes.Combining `was fine` with the temporal/state aspect:- states and their timing- detecting temporal statesConsider the most salient phrase: "was fine".Also

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

    languages and scripts

    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
    Not in Any Lists

    No Comments

    Negative Logits
    subsubsection
    0.49
    KS
    0.47
    delta
    0.47
     diverse
    0.47
    array
    0.46
    partisan
    0.45
    c
    0.45
     can
    0.45
    s
    0.45
     partisan
    0.44
    POSITIVE LOGITS
     вышла
    0.55
     впоследствии
    0.53
     πρώ
    0.52
    กับ
    0.51
     уже
    0.50
     первым
    0.49
     smlou
    0.49
     později
    0.49
     هاش
    0.49
     предложения
    0.48
    Activations Density 0.000%

    No Known Activations