© 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 EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-27B-IT
    3. 3-GEMMASCOPE-2-TRANSCODER-262K
    4. 87225
    Prev
    Next
    INDEX
    Explanations

    **analyze**`:The `MAX_ACTIVATING_TOKENS` list shows a strong presence of "ard" and "unate".The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list shows "ment" appearing multiple times.Combining these observations:- "ard" is a common English suffix (e.g., "hazard", "bastard", "wizard").- "unate" is also a suffix (e.g., "fortunate", "commiserate", "passionate").- When "unate" is there, "ment" often follows (e.g., "fortunate" -> "fortunately", "commiserate" -> "commiseration").The `TOP_POSITIVE_LOGITS` list includes "ING" and "ing". This suggests the neuron might be interested in words that can be turned into gerunds or participles, or words that are related to actions or ongoing processes.The `TOP_ACTIVATING_TEXTS` list contains:- "Quanti[hi]" (not very helpful)- "evolutionary robotics and artificial intelligence"- "spiritual power connected"- "Shogunate" (appears multiple times)- "Fuke sect"- "Tokugawa Shogunate" (appears multiple times)- "Ion Bombardment" (appears multiple times)- "Momentum Transfer"- "Tungsten with a small amount of thorium"- "vacuum tube"- "Edo Period"- "Jean-Luc Godard" -> "Godard" ends in "ard".- "Brownian motion" -> "Bombardment" ends in "ment", "motion" ends in "tion".Looking at `TOP_ACTIVATING_TEXTS`, "Shogunate" and "Tokugawa Shogunate" are prominent. "Shogunate" itself doesn't fit the "ard" or "unate" pattern directly, but it's a specific historical term."Godard" fits the "ard" pattern."Bombardment" ends in "ment", which follows "unate".The neuron seems to be picking up on specific word endings or parts of words, especially "ard", "unate", and potentially suffixes related to actions or states ("ment", "ing").Let's re-evaluate `MAX_ACTIVATING_TOKENS`: ard, unate, unate, unate, ard, ard, unate, arde, ard, ard.These are clearly parts of words."ard" is prominent."unate" is prominent."arde" is a variation.Let's look at the Top Positive Logits again: ING, utilisateur, ing, in, TERS, vær, culturale, illes, RO, grupi."ING" and "ing" are direct matches for gerunds/participles.The most direct pattern is the presence of "ard" and "unate" in the activating tokens. The subsequent tokens "ment" reinforce the "unate" pattern.Let's consider specific examples:- "ard" could be from "Godard".- "unate" could be from "fortunate", and the next token is "ment" (as in "fortunately" or "commiseration" after "commiserate").- "ard" could also relate to words like "hazard", "guard", etc.The `TOP_ACTIVATING_TEXTS` provide context:- Shogunate: historical term, ends in "-ate".- Godard: proper noun, ends in "-ard".- Bombardment: noun, ends in "-ment".- Momentum Transfer: noun phrase.The neuron seems sensitive to endings like "-ard" and "-ate/-unate", and potentially related suffixes like "-ment" and "-ing".Let's try to find a common thread.- "ard" endings: Godard, hazard, wizard.- "unate" endings: fortunate, commiserate.- "ment" endings: bombardment, movement, commiseration.- "ing" endings: evolutionary, spiritual, connected, ruling.The prompt asks for what the neuron "detects or predicts by finding patterns in lists".The pattern in `MAX_ACTIVATING_TOKENS` is strong for `ard` and `unate`.`TOKENS_AFTER_MAX_ACTIVATING_TOKEN` reinforces `unate` with `ment`.`TOP_POSITIVE_LOGITS` include `ing`."ard" and "unate" are word parts."ment" and "ing" are suffixes.Perhaps the neuron is detecting words with specific suffixes or word endings, or words that are common in certain contexts (like historical terms or technical terms).Let's re-read the request: "Keep your explanation concise (3 to 20 words). The explanation could be a single word, or phrase, or pattern. The explanation could be about tokens following or preceding certain tokens. The explanation could be about words starting with a sequence. Avoid simply listing all the tokens. Instead, try to find patterns."The most prominent *pattern* from `MAX_ACTIVATING_TOKENS` is the presence of `ard` and `unate` as word fragments.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` strongly suggests `ment` following `unate`.The `TOP_POSITIVE_LOGITS` suggest `ing` as a related concept.Let's try to combine these.- "ard" (e.g., Godard)- "unate" + "ment" (e.g., fortunate, commiseration)- "ing" (e.g., evolutionary, spiritual)This suggests a neuron interested in words indicating states, actions, or people, often characterized by specific endings/suffixes.Consider `ard` and `unate` as the primary "detector" part.The neuron detects words containing fragments like "ard" and "unate".The explanation should be concise.What if we focus on the *types* of words/concepts?- "ard" often refers to a person or thing (wizard, coward, postcard).- "unate" often refers to a state or quality (fortunate, passionate).- "ment" often refers to an action or result (punishment, achievement).- "ing" often refers to an ongoing action (running, singing).The neuron seems to activate for words that describe states, conditions, or ongoing processes, often signaled by "-ard", "-ate", "-ment", or "-ing" endings.Let's try to simplify. The most direct observed patterns *within* the provided lists are:1. `ard`2. `unate`3. `unate` followed by `ment`4. `ing` (from logits)These are all related to word endings or word fragments.Can we generalize this?Perhaps the neuron is looking for words that form concepts related to:- Personal attributes/descriptors ("ard", "unate")- Processes/states ("ment", "ing")Let's look at the activating texts for examples matching these.- "Shogunate" (historical term, ends "-

    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_3_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
    ("
    0.75
    mn
    0.71
    Ѡ
    0.66
     $-
    0.64
    new
    0.63
    댜
    0.63
    mvn
    0.63
    A
    0.63
    -
    0.62
    lz
    0.62
    POSITIVE LOGITS
    ING
    0.77
     utilisateur
    0.69
    ing
    0.64
     in
    0.61
    TERS
    0.61
     vær
    0.61
     culturale
    0.59
    illes
    0.58
    RO
    0.58
     групи
    0.58
    Activations Density 0.000%

    No Known Activations