HotpotQA

HotpotQA is a multi-hop question answering dataset built over English Wikipedia, where each question requires finding and reasoning over multiple supporting documents to answer. It provides sentence-level supporting facts (article title + sentence id) as strong supervision and to enable explainable predictions, and it introduces factoid comparison questions in addition to bridge-type questions. On Hugging Face it is offered in two configurations: 'distractor' (each question paired with 10 paragraphs: 2 gold + 8 Wikipedia distractors) and 'fullwiki' (open-domain, retrieve from all of Wikipedia). The distractor config has 90,447 train and 7,405 validation examples; the fullwiki config adds a 7,405-example test split. Questions are labeled by type (bridge/comparison) and difficulty level (easy/medium/hard).