"DeepVis Grammar of Ornament Multimodal Document Understanding Benchmark: Stratified Queries and Ground Truth Annotations"

"This dataset supports the evaluation of multimodal document understanding systems against a retrieval-augmented generation baseline on a 612-page visually rich historical corpus, Owen Jones's The Grammar of Ornament (1856). It contains 15 stratified evaluation queries across three categories \u2014 text-based (T), visual (V), and cross-modal (M) \u2014 each accompanied by annotated ground truth establishing correct page citations and expected content answers established prior to system evaluation. Full system responses from DeepVis and a RAG baseline using Azure AI Search with GPT-4.1 are included for all 15 queries. The dataset supports four evaluation metrics: Query Success Rate (QSR), Page Citation Precision (PCP), Page Citation Recall (PCR), and Visual Content Accuracy (VCA). It is intended for researchers evaluating vision-language systems, retrieval architectures, and citation accuracy on document collections where meaning depends on visual and spatial content rather than text alone."