Cohere Labs preprint finds a ‘culture funnel’ in LLM pipelines and publishes CultureMarkers dataset
Cohere Labs analyzed over 5.6 million training samples across pretraining, SFT, alignment and reasoning datasets and reports a consistent pattern — a ‘culture funnel’ where cultural diversity narrows as data moves into post-training stages. The team used Cohere’s Command A model to tag cultural signals, argues that multilingual coverage alone doesn’t ensure cultural representation, and published a preprint on arXiv plus the CultureMarkers dataset on Hugging Face to support further study.