LLM Training Data Cultural Representation

Cohere's research into LLM training data reveals a 'culture funnel' where cultural diversity significantly narrows from pre-training to post-training datasets. This analysis highlights that scaling multilingual capabilities alone does not guarantee culturally rich representation. The findings emphasize the need for intentional curation of culturally representative content within training data to build more globally inclusive AI systems. This iteration introduces a framework for surfacing and examining cultural signals at scale within training data, analyzing over 5.6 million training data samples across different stages of the LLM training pipeline and finding a consistent pattern: cultural diversity narrows as data moves from pretraining into post-training. The research also found that while adding more languages increases geographic diversity, it has diminishing returns in increasing the overall percentage of culturally grounded data, and that geographic representation is heavily skewed in cultural data. The analysis also shows that translation, local information requests, and message writing contain the strongest cultural signals, and that cultural awareness appears relevant across a wider range of tasks than current training distributions reflect.