© 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 EvalsExportsinterp-engineNEWAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-27B-IT
    3. 15-GEMMASCOPE-2-TRANSCODER-262K
    4. 14903
    Prev
    Next
    INDEX
    Explanations

    'django-auditlog''django.contrib.admin', 'django.contrib.auth', 'django'cljs.core.async :as async :refer [<! go chan put!]]''broadcast', exchange='broadcast', exchange_type='fanout'),''chat", help="Conversation style (chat or instruct)"'"line-to-neutral voltage."'hello file' content.")'cuda", help="Device to use (cuda or cpu)"'"model_size", default="base",'The `MAX_ACTIVATING_TOKENS` list contains snippets like `='`, `("`, `('`.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list contains many Python/programming related terms: `audit`, `django`, `cl`, `broadcast`, `chat`, `cuda`, `base`.The `TOP_ACTIVATING_TEXTS` show these tokens embedded within Python code snippets, configuration files, or command-line arguments, often defining modules, parameters, or lists. Examples include `'auditlog'`, `'django.contrib.admin'`, `'cljs.core.async'`, `broadcast`, `chat`, `cuda`, `base`.The pattern seems to be identifying specific keywords or identifiers that are often found in programming contexts, particularly within lists, assignments, or parameter definitions.Let's look for a more specific pattern:- `'auditlog'`- `django` (appears multiple times)- `cl` (appears in `cljs.core.async`)- `broadcast`- `chat`- `cuda`- `base` (appears in `model_size`)These look like names of libraries, modules, or specific configurations. The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are precisely these kinds of names. The `MAX_ACTIVATING_TOKENS` like `='` and `('` are often used when *assigning* these names or defining them within lists.Consider the `TOP_POSITIVE_LOGITS`: `around`, `written`, `Typed`, `VertexBuffer`, `Fach`, `Aerial`, `about`, `writing`, `Puzzle`, `jakarta`. This list is less clear. However, if we focus on the more concrete lists, the pattern emerges.The phrase "django.contrib.admin" suggests specific module names. "chat" and "broadcast" are also specific types. "cuda" is a specific device. "base" is a specific model size.The context seems to be picking out specific names, identifiers, or enumeration-like values, often found in configuration or code.Let's try to consolidate this. The neuron fires when it sees assignments or introductions to specific named entities within programming or configuration contexts.What if the neuron is about identifying *specific names or configurations* used in code/text?Example:- `'auditlog'` -> specific module name- `django` -> framework name- `cl` -> part of a library name (`cljs.core.async`)- `broadcast` -> type of exchange- `chat` -> type of style- `cuda` -> device name- `base` -> model size nameThe pattern is identifying *specific identifiers* used in lists, arguments, or definitions.A very common pattern in neural nets is identifying literal strings or specific keywords.specific module names

    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_15_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
     Date
    0.55
    ìn
    0.54
     fecha
    0.51
     date
    0.50
     wastes
    0.50
     lhs
    0.48
     Fecha
    0.46
     bá
    0.46
     evolução
    0.46
     lahko
    0.45
    POSITIVE LOGITS
    around
    0.52
    written
    0.50
    Typed
    0.49
    VertexBuffer
    0.46
    Fach
    0.46
    Aerial
    0.46
    about
    0.45
    writing
    0.45
    Puzzle
    0.45
    jakarta
    0.45
    Activations Density 0.000%

    No Known Activations