Cultural Awareness in Global AI | Cohere

This post details research into the 'culture funnel' in LLM training data, where cultural diversity narrows from pre-training to post-training. It introduces a framework for surfacing and examining cultural signals at scale within training data, analyzing over 5.6 million samples. Key findings include: 1) Cultural signals decrease across the training pipeline, with post-training data losing substantial diversity, often focusing on technical tasks. 2) Scaling multilingual capabilities increases geographic diversity but has diminishing returns on the percentage of culturally grounded data, with curation strategies being more impactful. 3) Geographic representation in cultural data is heavily skewed, with certain regions dominating. The post also highlights that translation, local information requests, and message writing tasks contain the strongest cultural signals, and that cultural awareness is relevant across more tasks than currently reflected in training distributions.