There is no posttraining data commons

Pretraining large language models benefited from a vast publicly available data commons—including Wikipedia, ArXiv, Common Crawl, and other open digital resources—that provided the raw material for the LLM revolution. However, posttraining (the next frontier in AI research) requires task-based data rather than documents, and no equivalent public commons exists for this purpose, creating a challenge for developing truly open advanced AI systems.