Datasets:
OCR evaluation references
catalog.json is a list of image records; images/ contains the corresponding files. Paths are relative to this directory. The initial selection contains 1,000 images.
Each record contains a stable id, image path, exact reference text, source category, reference scope, annotation quality, reading_order_known, optional word regions, and source provenance/rights. Source references and word regions are evaluation annotations, never hints to the OCR engine.
Keep source transcripts exactly as supplied. target_text_only references cover a selected target instruction, not necessarily all text visible in the image. The initial references have silver quality; preserve that distinction when reporting results. Dataset records and image bytes should be reviewed together.
The loading, metrics and runner code lives under research/ocr_eval/evaluation/ in the VIPI code repository.