© 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. 46-GEMMASCOPE-2-TRANSCODER-262K
    4. 139219
    Prev
    Next
    INDEX
    Explanations

    unique or specific words.However, the instructions state not to use phrases like "words related to".If I have to pick one commonality in TOP_POSITIVE_LOGITS: 'Special' appears twice in variations. 'Dried' and 'drying' are related. The others are quite disparate.Let's review the *behavior* of the neuron. It’s activating on *diverse* inputs. The highest contributing token list (MAX_ACTIVATING_TOKENS) is also very diverse (Chinese, English, numbers, punctuation). This suggests the neuron might be a general-purpose feature detector or one that reacts to the *presence* of specific, perhaps less common, features.Let's consider the possibility that the neuron is designed to detect specific *named entities* or *technical terms*, given the mix of names (McCartney, Paul), food items (辣), transport (地铁), and technical terms (TestMethod, Power SW).Given the instruction to be concise, and the failure of all previous methods, Method 4 is the last resort. The diversity of TOP_ACTIVATING_TEXTS makes it hard to pinpoint a single theme. However, let's consider the TOP_POSITIVE_LOGITS again: "TestMethod", "SPECIAL", "Dried", "Lim", "Sonder", "Specialized", "Die", "Special", "ioid", "drying". These are all English words. This suggests a potential focus on English, but the MAX_ACTIVATING_TOKENS include Chinese.Let's revisit Method 3. The words in TOP_POSITIVE_LOGITS are very varied.- TestMethod, Specialized, Special,TestMethod: Related to testing, specificity.- Dried, drying: Related to a state or process.- Lim: Name or part of word.- Sonder: Specific philosophical term.- Die: General word.- ioid: Suffix, likely technical.If I have to find *any* pattern here, it might be related to specific or technical terms.Let's try to simplify the explanation based on the logits. "Special", "Specialized", "TestMethod". These point towards some kind of specific designation or testing.Let's reconsider the entire set.MAX_ACTIVATING_TOKENS: 地铁, 辣, 低, work, no, Power, Paul, 風, 小, RedTOKENS_AFTER_MAX_ACTIVATING_TOKEN: 8, 粉, 落, permit, -, SW, McCartney, 破, 店, HouseTOP_POSITIVE_LOGITS: TestMethod, SPECIAL, Dried, Lim, Sonder, Specialized, Die, Special, ioid, dryingTOP_ACTIVATING_TEXTS: Mixed Chinese/English, mentions food, transport, names, technical terms.The most consistent aspect across the *logits* is the presence of specific, sometimes technical or abstract, English words. The highest logit is "TestMethod". "SPECIAL" and "Specialized" are also present. This suggests the neuron is highly predictive of specific or technical English terms.Considering the rules:- Concise explanation (1-6 words)- Do not add unnecessary phrases like "words related to"- Do not mention "tokens" or "patterns"- Specific explanationGiven the

    np_max-act-logits · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-it/transcoder_all/layer_46_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
     주제
    0.41
     Marquess
    0.39
    景观
    0.38
    憉
    0.38
    विकास
    0.38
    淐
    0.38
    飛
    0.37
     flew
    0.37
    亏
    0.37
    犃
    0.37
    POSITIVE LOGITS
    TestMethod
    0.42
     SPECIAL
    0.41
    Dried
    0.40
    Lim
    0.39
     Sonder
    0.38
     Specialized
    0.38
    Die
    0.37
    Special
    0.37
    ioid
    0.36
     drying
    0.36
    Activations Density 0.008%

    No Known Activations