© 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. 31-GEMMASCOPE-2-RES-262K
    4. 112301
    Prev
    Next
    INDEX
    Explanations

    ambient techno"},{"explanation": "Site Setup"},{"explanation": "escape route"},{"explanation": "only because"},{"explanation": "purring"},{"explanation": "buffer"},{"explanation": "vaccination site"},{"explanation": "hypothetical scenario"},{"explanation": "need it"},{"explanation": "Focus on"}The `MAX_ACTIVATING_TOKENS` list has `buffer`, `ambient`, `only`, `Site`, `need`, `the`, `escape`.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list has `techno`, `Setup`, `it`, `Focus`, `pur`.Let's look for patterns:- `ambient` is followed by `techno`. This suggests the neuron might be related to music genres like "ambient techno".- `Site` is followed by `Setup`. This suggests the neuron might be related to "Site Setup" or location setup.- `only` followed by `because` (contextually from text).- `escape` followed by `it` (contextually from text).- `need` followed by `it` (contextually from text).- `buffer` is often related to programming or data structures.The `TOP_POSITIVE_LOGITS` are a mix of various languages, which can sometimes indicate a neuron that is sensitive to specific words but not necessarily tied to a single language's domain in an obvious way.Considering the `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` and the `TOP_ACTIVATING_TEXTS`:- "atmospheric ambient techno" -> `ambient` + `techno`- "Vaccination Site Identification", "Site Setup" -> `Site` + `Setup`- "escape route…should you need it" -> `escape` + `need` (followed by `it`)- "Focus on" -> `Focus` (followed by `on`)The most prominent pairings seem to be "ambient techno" and "Site Setup". The presence of "escape" and "need" also points to specific scenarios.Let's try to find a unifying theme or the most common strong pattern."ambient techno" is a strong contender."Site Setup" is another.The explanations should be a single phrase or word. The prompt asks for what the neuron *detects or predicts*.Let's check the `MAX_ACTIVATING_TOKENS`: `buffer`, `ambient`, `only`, `Site`, `need`, `the`, `escape`.Let's check the `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `_`, `techno`, `because`, `Setup`, `it`, `*`, `Focus`, `pur`.Strong pairings:- `ambient` -> `techno` (music genre)- `Site` -> `Setup` (location/setup)- `escape` -> `it` or `need` (scenario/action)The `TOP_POSITIVE_LOGITS` are very diverse (Pashto, Turkish, Kannada, Hindi/Urdu, Korean, Tamil, Malay/Indonesian). This might suggest the neuron is sensitive to semantic concepts that appear across different languages rather than specific words in one language. However, the prompt says "finding patterns in lists".The instruction is "look at the lists and find a pattern that helps you explain the behavior of the neuron." and "The explanation should be specific."Let's re-evaluate the most frequent and specific patterns:1. `ambient` + `techno` (music)2. `Site` + `Setup` (location/organization)3. `escape` or `need` + `it` (scenario/action)"ambient techno" seems like a very common and specific phrase that arises."Site Setup" is also specific.If I have to pick one that shows a clear pairing from the given tokens, `ambient` followed by `techno` is a strong one. Also `Site` followed by `Setup`.The prompt asks for a concise explanation (3 to 20 words)."ambient techno" is 2 words."Site Setup" is 2 words.What if the neuron activates for descriptive phrases?Looking at the `TOP_ACTIVATING_TEXTS`:- "atmospheric ambient techno"- "Site Identification", "Site Setup"The `MAX_ACTIVATING_TOKENS` often pick up common words such as "the", "it", ".", etc. We should focus on the more semantically rich ones.`ambient` and `techno` pair directly.`Site` and `Setup` pair directly.`escape` and `need` appear in contexts related to scenarios.Is there a common theme between "ambient techno", "Site Setup", "escape route", "only because"?They seem to describe specific contexts or states.Let's consider "ambient techno". It's a specific genre.Let's consider "Site Setup". It's a specific operational step.Let's consider "escape route". It's a specific tactical element.The explanation needs to be a phrase.What if the neuron is about listing descriptive elements or states?The neuron may activate for specific descriptive phrases rather than general concepts.Let's consider the core pairings:- `ambient techno`- `Site Setup`- `escape route`- `need it`- `Focus

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-it/resid_post/layer_31_width_262k_l0_medium
    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
    1
    0.51
    一
    0.46
    率
    0.44
     wage
    0.44
    ঙ্গ
    0.43
     }
    0.42
    日
    0.41
     ausger
    0.41
     gas
    0.41
    ů
    0.41
    POSITIVE LOGITS
     وړیا
    0.54
    panelMenuList
    0.54
     gamanam
    0.53
     jaaye
    0.53
    dincer
    0.52
     پک
    0.52
    நே
    0.50
    ᕕ
    0.49
    acchati
    0.49
    tempat
    0.49
    Activations Density 0.004%

    No Known Activations