© 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-4-31B
    3. 30-RES-MATRYOSHKA-131K
    4. 130098
    Prev
    Next
    INDEX
    Explanations

    "between" appears in `MAX_ACTIVATING_TOKENS` and is followed by "the" in `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` in the `TOP_ACTIVATING_TEXTS`. The `TOP_POSITIVE_LOGITS` are diverse, suggesting the neuron might not be about a specific concept but rather structural or positional.However, looking deeper:1. `MAX_ACTIVATING_TOKENS`: `ha`, `between`, `Set`, `in`, `://`, `a`, `u`, `frac`, `"`, `from`2. `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `pregunt`, `the`, `env`, `seconds`, `Team`, `0`, `{`, `google`, `{`3. `TOP_POSITIVE_LOGITS`: `তাড়াতাড়ি`, `piel`, `paparazzi`, `䣼`, `கலங்கரை`, `सबसे`, `poświę`, `鿷`, `cuatro`, `circumcision`4. `TOP_ACTIVATING_TEXTS`: * "...no ha pregunt" (Spanish) -> `ha` followed by `pregunt` * "...happens between the Go JBL..." -> `between` followed by `the` * "...Set env variable..." -> `Set` followed by `env` * "...7/4 of a hour in seconds?" -> `frac` (implicitly via 7/4) and `seconds` * "...Booked a Team Party Venue" -> `a` followed by `Team` * "default_endpoints_protocol = std::env::var..." -> `env` follows a preceding token, `var` (not seen but inferred context), and `{` follows `google` * " $\binom {n} {i,\ j,\ n-i-j}$? Which one of the two is $\frac{n!}{i!$ -> `frac` (via n!) and implicitly a number/term that could be like 0.The pattern `[token] [specific word/number]` seems common.Specifically, pairs like `between the`, `Set env`, `a Team`, `frac seconds`.The `TOP_POSITIVE_LOGITS` being unrelated words in different languages make it hard to pinpoint a semantic meaning.Let's re-examine `MAX_ACTIVATING_TOKENS` with their subsequent tokens in mind.`ha` -> `pregunt` (Spanish 'asked')`between` -> `the` (English article)`Set` -> `env` (environment variable acronym)`in` -> (no clear pair in the provided truncated list)`://` -> (URL prefix, often followed by domain 'google')`a` -> `u` (possibly part of a word like 'a unique' or 'a user'. But `a` is followed by `Team` in text)`u` -> (no clear pair)`frac` -> `seconds` (mathematical fraction followed by a unit)`"` -> (quotes, often preceded/followed by words)`from` -> (preceding 'googlebot')The `TOP_POSITIVE_LOGITS` are very diverse, suggesting no single semantic concept is captured.The prompt emphasizes finding patterns in lists, especially tokens following max activating tokens.The pattern of a specific token being followed by another specific token seems strong.`between the`, `Set env`, `frac seconds`.Also, `ha pregunt`.This neuron seems to be about specific token sequences, often involving common technical or grammatical pairings, especially in programming contexts (`env`), mathematical contexts (`frac`, `seconds`), or descriptive contexts (`between the`).Let's look at the TOP_POSITIVE_LOGITS again. They are single words in various languages. This might indicate the neuron is sensitive to *language shifts* or *specific types of entities* that are multilingual, but that's too vague.Focusing on the sequential nature:`MAX_ACTIVATING_TOKENS`: `ha`, `between`, `Set`, `in`, `://`, `a`, `u`, `frac`, `"`, `from``TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `pregunt`, `the`, `env`, `seconds`, `Team`, `0`, `{`, `google`, `{`This neuron seems to trigger on common combinations or sequences found in text, particularly those that might bridge different types of information (e.g., a phrase followed by a unit, a command followed by a variable name, a preposition followed by an article).The "specific" aspect is key.`between the``Set env``frac seconds``ha pregunt`These are not abstract concepts. They are specific pairings.The explanation should be concise and capture *what* it detects or predicts via patterns.Consider the types of input:- Spanish sentence chunk- English pairing instructions- Rust code snippet- Math problem (unit conversion, fractions, probability)- Package version link- Customer feedback- Code constant declaration (`:`)- LaTeX math- Search engine variable setupThe `TOP_POSITIVE_LOGITS` are extremely diverse: Bengali, Spanish, English, Chinese, Tamil, Hindi, Polish, and others. This suggests the neuron is *not* about specific words or concepts, but more likely about *structure*, *syntax*, or *common transitional sequences*.The strongest pattern is `[token] [common following token]`.Let's try to generalize this."specific token sequences" is too broad."common word pairings" is better."token-following-token sequences" - mentions tokens.What if the specific token itself is the key, and what follows is the confirmation?`ha` -> `pregunt` (Spanish)`between` -> `the` (English)`Set` -> `env` (Programming)`frac` -> `seconds` (Math/Units)The neuron might be detecting phrases or sequences that *transition* between different types of information or common structures.From the `TOP_POSITIVE_LOGITS`: `cuatro` (four), `paparazzi`, `circumcision`From the `TOP_ACTIVATING_TEXTS`: `seven`, `four`, `hours`, `seconds`, `Team`, `google`, `env`, `pair`, `between`.The `TOP_POSITIVE_LOGITS` are too varied to extract a single semantic theme.The `MAX_ACTIVATING_TOKENS` are also quite varied.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are also varied.However, the *combination* of `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` shows specific, frequent pairings in the `TOP_ACTIVATING_TEXTS`.`between the``Set env``frac seconds` (implicitly, given

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    decoderesearch/gemma-4-saes/gemma-4-31b
    Prompts (Dashboard)
    16,384 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     和
    -0.06
     在
    -0.06
     médias
    -0.06
    还将
    -0.06
    however
    -0.05
     ayrıca
    -0.05
     -
    -0.05
    ,
    -0.05
    larını
    -0.05
    son
    -0.05
    POSITIVE LOGITS
     তাড়াতাড়ি
    0.06
     piel
    0.06
     paparazzi
    0.06
    䣼
    0.06
     கலங்கரை
    0.06
     सबसे
    0.06
     poświę
    0.06
    鿷
    0.06
     cuatro
    0.06
     circumcision
    0.06
    Activations Density 0.001%

    No Known Activations