© 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. 1-GEMMASCOPE-2-TRANSCODER-262K
    4. 108915
    Prev
    Next
    INDEX
    Explanations

    The neuron seems to be detecting chemical compounds, specifically flavonoids like kaempferol and quercetin, often associated with health and antioxidant properties.Let's refine this."emp" appears multiple times."ecal" appears multiple times."char" appears once.followed by:"fer" (e.g., kaempferol)"cul" (e.g., calculator)"лый" (Russian suffix)The TOP_POSITIVE_LOGITS have 'offer'. The TOP_ACTIVATING_TEXTS contain many chemical compounds, especially flavonoids ('quercetin', 'kaempferol', 'luteolin'). The text also mentions "tdeecalculator".The "emp" tokens are often followed by "fer" which is part of words like "kaempferol".The "ecal" tokens are often followed by "cul" which is part of "calculator".The Cyrillic 'с' is followed by 'лый' which seems to be a Russian ending.The most prominent pattern seems to be the detection of specific chemical compounds (flavonoids) and potentially terms related to energy calculation.Let's look for a common theme that links these.'kaempferol' is a flavonoid.'tdeecalculator' is for energy expenditure.Are the activating texts hinting at specific types of compounds or processes?Flavonoids, antioxidants, Phase 1 enzymes, folate, B Vitamins.Total Daily Energy Expenditure (TDEE) calculator.Domoic acid (algal toxin), Hepatitis A virus (fecal-associated virus), Norovirus (fecal-associated virus), Salmonella.This is quite diverse. However, 'kaempferol' is a strong signal from both MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN ('emp' -> 'fer').Looking at TOP_POSITIVE_LOGITS: 'offer' and 'you'.The phrase needs to be concise (3-20 words).It could be about specific words starting with a sequence, or tokens following a token.Consider the 'emp' followed by 'fer'. This strongly suggests words containing 'empfer'. 'kaempferol' is the prime example from the text.Also, 'ecal' followed by 'cul' suggests words containing 'ecalcul'. 'tdeecalculator' is an example.The phrases "words like kaempferol" or "energy calculation terms" are too descriptive and not just the pattern itself.What if the neuron detects specific chemical names or related scientific terms?The lists aren't strongly pointing to one thing.Let's re-evaluate the output lists.MAX_ACTIVATING_TOKENS: emp, emp, emp, ecal, emp, emp, с, ecal, ecal, charTOKENS_AFTER_MAX_ACTIVATING_TOKEN: fer, fer, fer, cul, fer, fer, лый, -, -, ):Pattern observed:`emp` followed by `fer` (suggests `kempferol`)`ecal` followed by `cul` (suggests `calculator`)`с` followed by `лый` (Russian text)TOP_POSITIVE_LOGITS: to, on, t, you, <0x0D>, On, </h2>, offer, </span>, yThis list is less informative for a specific pattern. 'offer' is in the text.TOP_ACTIVATING_TEXTS: Contain mentions of:- Flavonoids (quercetin, kaempferol, luteolin, apigenin, epigallocatechin gallate)- Vitamins (folate, B9)- Antioxidants- TDEE calculator- Toxins (Domoic acid, Brevetoxin)- Viruses (Hepatitis A, Norovirus)- Bacteria (Salmonella)The strongest, most consistent pattern across MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN is sequences that predict specific terms, like "kempferol" from "emp"+"fer", and "calculator" from "ecal"+"cul". However, one is a chemical compound, the other is a tool.The repetition of 'emp' + 'fer' is striking. This strongly points to 'kaempferol' and similar flavonoid names.The 'ecal' + 'cul' points to 'calculator'.If I have to choose one behavior, the chemical names seem more dominant in terms of quantity and word structure. The 'emp'+'fer' is a very specific partial word matching.Could the neuron be detecting specific chemical suffixes or prefixes?'emp' + 'fer' -> part of kaempferol'ecal' + 'cul' -> part of calculatorLet's consider the possibility that the neuron is sensitive to specific alphanumeric sequences that form parts of certain types of words.The `emp` token followed by `fer` and `ecal` followed by `cul` are strong signals for specific compound names (like `kaempferol`) and calculation tools (`<tdee>calculat<or>`).The prompt asks for "what the neuron detects or predicts by finding patterns in lists".The most specific pattern detected is the sequence of tokens that are word fragments.`emp` -> `fer``ecal` -> `cul`This suggests the neuron predicts specific scientific/technical terms.The `TOP_POSITIVE_LOGITS` contains `offer`, and `offer` appears in the texts, e.g. "offer health benefits".Considering 'kaempferol' is present in both activating texts and the token sequence:'emp' is part of kaempferol.'fer' follows 'emp'.This is a direct link.What if the neuron is detecting negative health-related substances or common substances in general?Flavonoids, Quercetin, Kaempferol are positive.Toxins, Viruses, Bacteria are negative.Calculator is neutral/tool.The neuron is likely detecting specific *names* or *types* of entities.Let's try to summarize the `emp`/`fer` and `ecal`/`cul` patterns:`emp` followed by `fer` predicts parts of chemical names.`ecal` followed by `cul` predicts parts of technical terms (like calculator).The phrase should not mention "tokens" or "patterns".How about focusing on the identified chemical compound 'kaempferol', as 'emp'+'fer' is a very strong indicator in the lists?If I had to pick *one* specific thing it detects, it would lean towards chemical compounds due to the repetition in MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN AND TOP_ACTIVATING_TEXTS.Let's consider the "specific chemical compound" angle.The neuron seems to detect chemical compounds and related scientific terms.Specifically, `emp` followed by `fer` points to `kaempferol`.The `TOP_POSITIVE_LOGITS` includes `offer`, and `offer` appears in the phrase "offer health benefits".The `TOP_ACTIVATING_TEXTS` feature many chemical compounds like quercetin and kaempferol, as well as tools like TDEE calculator, and biological hazards

    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_1_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
     في
    1.29
    ين
    1.26
    在
    1.25
     in
    1.11
    も
    1.02
     σε
    0.98
    <0x80>
    0.97
    ева
    0.97
    يلي
    0.96
    ні
    0.95
    POSITIVE LOGITS
     to
    1.90
     on
    1.63
    t
    1.52
     you
    1.39
    <0x0D>
    1.27
     On
    1.27
    </h2>
    1.22
    offer
    1.20
    </span>
    1.16
    y
    1.16
    Activations Density 0.000%

    No Known Activations