A 100k-row retrieval QA dataset where each row contains a question, ground-truth chunks, and pre-mined distractor chunks (random + semantically similar). Designed for training and evaluating retrieval ...