LLM-ready, augmented dataset with chunking/paraphrasing, back-translation, self-consistency, counterfactuals.