© 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. 204955
    Prev
    Next
    INDEX
    Explanations

    "1-6 words, mostly 1-3 words." "Do not add unnecessary phrases like 'words related to', 'concepts related to', or 'variations of the word'. Do not mention 'tokens' or 'patterns' in your explanation."Considering Method 3 again:TOP_POSITIVE_LOGITS: Kim, TextChanged, Kevin, oga, ukup, Kim, mnf, लाहि, সাপোর্ট, 김The most specific repeating element is "Kim".If Method 3 is chosen: "Kim".Let's consider Method 4 again, ignoring max activating tokens and looking at top activating texts for broad themes."Lean 4 code", "theorem", "Interactive Theorem Proving", "transcription".The strongest *contextual* theme appears to be related to programming languages like Lean and formal verification (theorem proving).If Method 4 is chosen: "Lean theorem proving". This is 3 words, fits the constraint.Which method takes precedence? "Once you find a pattern, stop and return that pattern. Do not proceed to the later methods."Method 3: Finds "Kim" as the most specific recurring element.Method 4: Finds "Lean theorem proving" as a theme.The prompt asks to find "what this neuron detects or predicts". "Kim" is a specific thing. "Lean theorem proving" is a context. The logits directly suggest "Kim".Let's assume Method 3 is preferred if it yields a specific item.Method 3: "Kim" is the most direct pattern in the TOP_POSITIVE_LOGITS.Let's consider the prompt's example for Method 2: 'say [the pattern]' if they all start with the same letter. This implies precision.Re-evaluation of Method 3:`TOP_POSITIVE_LOGITS`: Kim, TextChanged, Kevin, oga, ukup, Kim, mnf, लाहि, সাপোর্ট, 김The most striking feature is the repetition of 'Kim'. Is this sufficient for a pattern? Yes, repetition is a pattern. "Kim" is concise.If Method 3 is used, the explanation is "Kim".Let's double check if Method 4 could be *more* appropriate. Method 4 is a "best guess by describing the broad theme or context". Method 3 is "looking for similarities". Repetition of "Kim" is a similarity.If I have to pick the *most direct* pattern from the lists, Method 3 looking at logits is often very direct. The other items in logits might be supporting evidence for a broader theme if "Kim" alone felt too weak or coincidental. But the prompt says "If they share something specific in common". Repetition is a specific commonality.Let's lean towards Method 3.However, the example in the *user's* instruction set implies that multiple tokens under Method 3 could lead to a broader description like "fruits" in Neuron 2, not just a single repeating token. "Kim" is repeated once, which is a stronger signal than if it appeared once.Let's consider if the *combination* of things in logits points somewhere. Names (Kim, Kevin) + technical term (TextChanged) + foreign words. This feels less cohesive than the "fruits" example.Let's look at the texts again, ignoring the max activating tokens.Text 1: "Lean 4 code defining the fixed derivative of a polynomial"Text 2: "Interactive Theorem Proving", "LEAN verifies each step"Text 3: "medical transcription", "patient", "symptoms"Text 4: "transcription costs"

    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
     pre
    0.38
     Jokes
    0.37
     FOCUS
    0.37
     jokes
    0.36
    Recipes
    0.35
     परिचय
    0.34
     आरोग
    0.34
     heath
    0.34
    㷅
    0.34
     prescribes
    0.34
    POSITIVE LOGITS
    Kim
    0.44
    TextChanged
    0.41
    Kevin
    0.40
    oga
    0.39
    ukup
    0.39
     Kim
    0.38
    mnf
    0.37
     लाहि
    0.37
     সাপোর্ট
    0.37
    김
    0.37
    Activations Density 0.000%

    No Known Activations