© 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-IT
    3. 40-GEMMASCOPE-2-TRANSCODER-262K
    4. 32268
    Prev
    Next
    INDEX
    Explanations

    The presence of "support" and "compat" suggests a general concept related to aid or assistance. The other words are less clear or less consistently patterned.**TOP_ACTIVATING_TEXTS:**- "Ukraine is pleading with Western allies for more aid, particularly air defense systems and ammunition."- "Ukraine is pleading for more aid from Western allies."- "Ukraine is pleading with Western allies for more aid, especially air defense systems and artillery ammunition."- "Ukraine is pleading for continued Western aid."- "Ukraine is appealing for more aid from its allies."- "Ukraine is appealing for more aid from Western allies."- "Ukraine is appealing for continued Western military and financial aid from Western allies."- "Ukraine is appealing to the West for more heavy weaponry."- "Ukraine is urging allies for more air defense systems."The texts overwhelmingly focus on Ukraine requesting aid from Western allies, particularly in the context of war. The verbs used are "pleading", "appealing", and "urging".**Synthesizing the pattern:**The `MAX_ACTIVATING_TOKENS` are overwhelmingly "pleading", "appealing", and "urging".The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` frequently include "for" and "allies", indicating the nature of the request.The `TOP_ACTIVATING_TEXTS` confirm this pattern: Ukraine asking Western allies for aid, especially military aid.`TOP_POSITIVE_LOGITS` are a bit more mixed, but "support" points to the overarching theme of aid/assistance.The core pattern is a strong request (pleading/appealing/urging) for aid from allies.**Constraint Checklist & Confidence Score:**1. Concise explanation (3-20 words): Yes2. Captures what neuron detects/predicts: Yes3. Finds patterns in lists: Yes4. Not just listing tokens: Yes5. No "words related to" etc.: Yes6. No "This neuron detects/predicts": Yes7. No "tokens" or "patterns": Yes8. No capitalization unless proper noun: Yes9. Specific explanation: Yes10. Majority match pattern: Yes11. If cannot guess, use first MAX_ACTIVATING_TOKEN: N/A (can guess)Confidence Score: 5/5**Mental Sandbox:**- "pleading for aid" - 3 words, specific. Fits text.- "appealing for support from allies" - 6 words, specific. Fits text.- "calling for aid from allies" - 5 words. "Calling for" isn't directly in MAX_ACTIVATING_TOKENS, but "pleading/appealing/urging" are synonyms.- "Ukraine requesting aid from allies" - 6 words. "Ukraine" is specific to the text, but the neuron might be more general. However, let's check if the neuron is only activated by *Ukraine* asking. The examples are mostly about Ukraine specifically.- "making strong requests for aid" - 5 words. Good, captures the essence.- "pleading/appealing/urging for allies' aid" - 6 words. This is very accurate.Given the strong presence of "pleading", "appealing", and "urging" in `MAX_ACTIVATING_TOKENS`, and the consistent theme of requesting aid from allies in `TOP_ACTIVATING_TEXTS`, a phrase combining these is best.Let's try "pleading to allies for aid". This uses a verb from MAX_ACTIVATING_TOKENS, and tokens from TOKENS_AFTER_MAX_ACTIVATING_TOKEN ("for", "allies"), and captures the core meaning of the texts. It's 5 words.Another option: "appeals for support from allies". This captures the alternative verb "appealing" and the concept of "support" from TOP_POSITIVE_LOGITS and the text.Let's re-evaluate the `TOP_POSITIVE_LOGITS`. "support" is one of the top ones. "pleading" and "appealing" are clearly requests for support/aid.How about something that encapsulates the *type* of request? It's often about "aid", "systems", "ammunition", "heavy weaponry". But "aid" is a good generalization."pleading for aid from allies" - 6 words. Very specific and accurate."urging allies for support" - 4 words. Also good.Looking at the specific text examples, "pleading with Western allies for more aid" is a very common phrase."pleading with Western allies for more aid" -> becomes "pleading ... for ... allies""Ukraine is pleading with Western allies for more aid"The neuron seems to activate when "pleading", "appealing", or "urging" are followed by prepositions ("for", "with") and entities like "allies" or similar groups, in the context of requesting assistance.Let's try "pleading/appealing to allies for aid". This covers the verbs and the goal.Or simply "pleading for aid from allies".Let's consider the most direct capture of tokens.MAX_ACTIVATING_TOKENS: pleading, appealing, urgingTOKENS_AFTER_MAX_ACTIVATING_TOKEN: for, with, to, alliesA simple combination would be:"pleading for aid" (3 words) - might be too general."pleading to allies" (3 words) - good, but misses the 'what'."appealing for aid" (3 words) - good.The texts are *all* about Ukraine requesting aid.Let's aim for specificity based on the text."asking allies for aid" - 4 words. "asking" isn't in MAX_ACTIVATING_TOKENS, but implies the verb."pleading for aid from allies" - 6 words. This seems very robust.Let's check the prompt's preference for specificity. "unique words" is not specific enough. "foreign words" is not specific enough. "pleading for aid from allies" is specific.How about focusing on just the action and the receiving end?"pleading with allies" - 3 words."appealing for support" - 3 words.The fact that "support" is high in logits and "aid" is prevalent in texts suggests that the output might be related to "support" or "aid".Let's try: "pleading for support from allies" (5 words) or "appealing for aid from allies" (5 words)."appealing for aid from allies" feels slightly more direct and covers the most common pattern.The prompt asks for the *neuron's behavior*. The neuron is activated by the *action* of pleading/appealing/urging.Final check: "appealing for aid from allies"- Conciseness: 5 words (fits 3-

    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_40_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
     зміню
    0.47
     ضد
    0.46
    against
    0.45
    Against
    0.45
     against
    0.43
     מא
    0.42
    對
    0.42
     والمع
    0.41
     tegen
    0.41
     зона
    0.41
    POSITIVE LOGITS
     поддер
    0.40
     pesar
    0.37
    یدم
    0.37
     Vivo
    0.37
     compat
    0.36
     viv
    0.36
     support
    0.36
     awak
    0.35
     poroz
    0.35
     karar
    0.35
    Activations Density 0.001%

    No Known Activations