A Curated Logistics Document Dataset for Layout-Informed Key-Value Extraction

This dataset is a curated benchmark comprising 200 real-world logistics documents designed to support research in document semantic layout analysis and key-value information extraction. It specifically addresses the challenges of complex, diverse, and template-variant forms common in logistics and supply chain operations. The dataset is structured for a modular Document AI pipeline, offering three distinct components: (1) Raw PDF documents, (2) Visual layout annotations in YOLOv8 format (TXT and JPG images) for training object detection models to identify logical zones, and (3) JSON annotations for key-value extraction (semantic ground truth). This resource was integral to the development and evaluation of the Layout-Informed Chunking (LIC) pipeline described in the associated IEEE Access manuscript, providing a crucial resource for researchers aiming to replicate and advance modular Vision-Language model approaches in Document AI.