© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    Natural Language
    Autoencoders
    NEW
    Assistant AxisNEWCircuit TracerUPDATESteerSAE EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-27B-IT
    3. 31-GEMMASCOPE-2-TRANSCODER-262K
    4. 51484
    Prev
    Next
    INDEX
    Explanations

    decimal expansion of 1/7? (This requires finding the repeating pattern). (Difficulty: 3)- "What is the compound interest accrued on a sum of ₹8000 at the rate of 10%"- "NCERT math 'Quadratic Equations'.mathematics questions about linear equation and with answers?- "Okay, let's prove the **Distributive Property**. This is a fundamental property in Algebra 1 and"- "Okay, here are three math questions about linear equations, with answers and explanations. I've tried to"- "Yes, absolutely! I can definitely try. Please give me the problems you'd like me to solve."- "Can you solve some mathematics problems?"- "Resistor * Variable Resistor (Rheostat) * Ammeter * Voltmeter"- "Switch (Open and Closed) * "- "Okay, here's a presentation outline covering the topics you"- "Derivatives 2. Homogeneous functions 3. Euler's theorem 4. Tangents & Normal 5. Asymptotes 6. Jacobians"- "a polynomial) - just substitute the value. * *Example:* lim (x2 + 3x - 1) as x -> 1 = (1)2 + 3(1)"- "Prime Time: Find the smallest prime number greater than 100 that is the sum of two other prime"- "2. Quantitative Aptitude (Easy) What is the compound interest accrued on a sum of ₹8000 at the rate of 10%"- ") 360 meters (e) 216 meters **2. Quantitative Aptitude (Easy)"- "Polynomials"**Analysis:**- `MAX_ACTIVATING_TOKENS` contains punctuation and structural tokens like `<start_of_turn>`. These are often less indicative of semantic meaning than other tokens.- `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` has "model" repeatedly. This suggests the neuron is activated when the model is outputting text, which is expected. However, it also has "ympt" which could relate to "asymptotes" seen in the text.- `TOP_POSITIVE_LOGITS` contains a mix of languages and terms ('Digital', 'graphs', 'Leave', 'காலமாக', 'დილ', 'მწიფ').- `TOP_ACTIVATING_TEXTS` is the most informative. It heavily features mathematical concepts: - Numbers, HCF, LCM, decimal expansions, rational numbers, prime numbers. - Polynomials, quadratic equations, linear equations. - Compound interest, quantitative aptitude. - Derivatives, asymptotes, Jacobians. - Geometric concepts (meters). - Mathematical properties (Distributive Property). - The phrase "mathematics problems" also appears.**Pattern Identification:**The overwhelming theme across `TOP_ACTIVATING_TEXTS` and some elements in `TOP_POSITIVE_LOGITS` (like 'Digital', 'ographiques' implying visual/graphical representation of math) and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` ('ympt' from asymptotes) points towards **mathematics**.The specific items include:- "decimal expansion"- "prime number"- "compound interest"- "linear equation"- "Distributive Property"- "Asymptotes"- "Mathematics problems"All these fall under the umbrella of mathematical topics or questions. The neuron seems to activate when the model is discussing or generating mathematical problems, concepts, or solutions.**Concise Explanation:**Looking for a pattern that integrates these:- "mathematics problems" (appears directly)- "decimal expansion", "prime number", "compound interest", "linear equation", "Distributive Property", "Asymptotes" are all specific math topics.The neuron is identifying when the subject matter is `mathematics`.Let's refine the phrase to be specific and within the word count."mathematics problems" is a direct match for a phrase in the data and covers many specific examples."mathematical topics" is more general."equations and numbers" covers a subset.Given the variety within mathematics, "mathematics problems" or "math problems" seems appropriate and specific according to the examples.Let's consider the other options.- "decimal expansion" - too specific, only one example.- "prime number" - too specific.- "asymptotes" - too specific.- "equations" - too broad, only one type mentioned explicitly."mathematics problems" is a strong candidate."math problem generation" could also work.Reviewing the rules:- Concise (3 to 20 words)- Single word or phrase- About tokens following/preceding, words starting, etc.- Avoid listing all tokens- Specific- Not "This neuron detects/predicts"- No "tokens" or "patterns"- No capitalization unless proper noun"mathematics problems" is a good candidate. It directly appears in the text and encapsulates the type of content the neuron is activating on.Consider other positive logits: 'Digital', 'Siz', 'Leave', 'Dire', 'oids', 'ographiques'.These are broader and might indicate something about structured text or specific types of concepts, but the `TOP_ACTIVATING_TEXTS` is much stronger evidence for math.So, the core is mathematics, specifically problems or examples."mathematics problems" fits."math problem generation" also fits."computational math topics" is also an option.Let's re-examine `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: "ympt", "model", "model", "model"."ympt" strongly suggests "asymptotes".The repetition of "model" is likely just the model's own output marker.The phrase "mathematics problems" seems to be the most direct and accurate summary of the observed activations and logits. It's specific enough (identifying the domain) and covers the types of entries seen.Final check:- "mathematics problems" is 2 words. (3-20 words rule - it's short, maybe needs more?)- "finding patterns in lists" - my task.The explanation should capture *what the neuron detects or predicts*.The neuron is detecting/predicting the presence of `mathematics problems` or `math content`.What if we combine some elements?"math equations and number theory" - covers some, but not all."mathematical concepts and questions" - more general again.Let's see the examples again:- "Real Numbers: Questions on the HCF and LCM of numbers, and determining if a rational number has a terminating or non-terminating decimal expansion."- "What is the compound interest accrued on a sum of ₹8000 at the rate of 10%"- "NCERT math 'Quadratic Equations'."- "mathematics questions about linear equation"- "Distributive Property"- "Derivatives", "Asymptotes"The common thread is `mathematics`. The presence of "Questions", "Problems", "Equations", "Numbers", "Properties", "Concepts" points to asking for or providing mathematical tasks or information.Consider "mathematical problems and equations

    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_31_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
    使い方
    0.42
     поверхность
    0.36
     direito
    0.35
     superfície
    0.35
     utilización
    0.35
     labores
    0.35
     الهند
    0.34
     meticulous
    0.34
     કાર્ય
    0.33
     pavimento
    0.33
    POSITIVE LOGITS
    Digital
    0.39
     Siz
    0.39
     காலமாக
    0.37
    Leave
    0.37
    Dire
    0.36
    დილ
    0.36
    მწიფ
    0.35
     Leave
    0.35
    oids
    0.35
    ographiques
    0.34
    Activations Density 0.000%

    No Known Activations