Herder in the lab: Maximizing “cultural density” in multilingual training data for an open‐science AI project

This paper provides ethnographic analysis of interactions between volunteers in an open‐science AI project entrusted with preparing multilingual text for use in fine‐tuning a pretrained LLM. Volunteers were encouraged to entextualize translated text into structured corpora and to write and edit prompt‐completion pairings in over 100 languages. This drive to capture linguistic diversity often conflicted with the need to organize data sets into standardized categories to effectively fine‐tune the LLM. Volunteers purified text annotations by judging what constituted normatively correct syntax, punctuation, spelling, and what kinds of content best represented the “essence” of their culture. Certain forms of language were described as “culturally dense,” suggesting an underlying linguistic ideology aligning with Herder's Romantic notions of cultural relativism.