Multi Hop Reasoning Dataset Construction With Supporting Fact Annotation

via “multi-hop reasoning dataset construction with supporting fact annotation”

113K questions requiring multi-hop reasoning across Wikipedia articles.

Unique: Explicitly annotates supporting facts at sentence-level granularity rather than just providing QA pairs, enabling evaluation of both answer correctness AND reasoning transparency. The dataset design enforces multi-hop requirements through crowdsourcing validation that questions cannot be answered from single documents.

vs others: Differs from SQuAD (single-document QA) and MS MARCO (web-scale but less structured) by providing explicit multi-hop reasoning requirements with supporting fact labels, making it uniquely suited for training interpretable reasoning systems rather than just answer extraction.