PURE-MMQA: A Task-Aware, Provenance-Preserving Multimodal Dataset for Question Answering over Software Requirements Specifications

PURE-MMQA is a task-aware and provenance-preserving multimodal dataset for question answering over Software Requirements Specifications (SRSs). The dataset contains text-grounded and figure-grounded QA items generated from a controlled subset of the PURE requirements corpus. Each QA instance includes task labels, modality information, page references, section hints, direct answers, generated answers, and multi-judge evaluation scores to support auditability and retrieval-augmented generation evaluation.