structured data format delimiters and metadata field names in JSON, XML, and YAML files.

refusal language and the opening statements of safety guardrail responses that reject harmful requests.

formatted explanations with bullet points and structured sections.

phrases where the model acknowledges a user's emotional distress, refuses a harmful request, and offers supportive alternatives or resources.

tokens related to journalism, news sources, information dissemination, and detailed reporting.

negation or correction structures that clarify what something is not before explaining what it actually is.

code execution and output statements, particularly print functions, function calls, and test assertions.

tokens related to errors, warnings, harmful content, deception, and security threats.

words describing physical substances, materials, ingredients, and their properties (such as texture, state, and composition).

references to geographic locations, countries, regions, and expressions of global or international scope.

descriptive and persuasive language that conveys instructional intent, emotional significance, or substantive meaning.

foreign-language words and names, especially those with diacritics, non-Latin scripts, or hyphenated reduplication.

references to people or collective human groups, often in constructions with relative clauses referring to them.

references to clandestine, mission-oriented actions—such as heists, assassinations, and tactical operations—covering planning, execution, and the agents involved.

meta-discursive, emphatic framing in prose—generalized statements and reaction/argument setup using intensifiers, quantifiers, and function-word-heavy constructions.

The neuron primarily detects numeric tokens (digits, numerals and year-like numbers) in the text.

tokens involved in the model's self-introduction—first-person "I" + the contraction "m" and the assistant's name/identity.

mentions of the model's identity, creators, and related branding/attribution tokens (e.g., Gemma, team, created, Google, parts of the AI URL).

promotional marketing language that includes calls to action or persuasive messaging.

Detects the lexical token for thinking/cognition (the verb and its appearances in multiword phrases and compounds).