This dataset packages a frozen, publication-ready benchmark for dense visual document editing: 3,506 examples spanning text and table edits in Chinese and English, with embedded image bytes, allowed-edit masks, actual-difference masks, and structured JSON sidecars for provenance and annotations. It is intended to enable repeatable model evaluation at the OCR-parsing level rather than act as a source of training-only synthetic images.
What Sets It Apart
- OCR-grounded evaluation: examples include OCR/annotation sidecars and region-level targets to support per-block matching, BLEU/TEDS-style metrics and precise edit-region assessment. This makes the collection useful for measuring text-preservation, style consistency, and localized editing accuracy rather than only perceptual image quality.
- Rich, dual-mode coverage: the packaged splits and configs expose text-only, table-only, and combined subsets, plus compatibility assets (packed tar archives) and Parquet rows with embedded image/mask bytes for convenient loading in data pipelines.
- Provenance and sanitization preserved: source provenance, asset manifests, and source-specific license notes are retained; publication metadata is sanitized (no annotator personal metadata) and the snapshot is frozen with documented build dates.
Who it's for and trade-offs
Great fit if you need a benchmark to evaluate or compare image-editing models on dense, bilingual document pages—especially for research targeting OCR-aware edit accuracy or table-structure edits. Look elsewhere if you require a large-scale, fully permissively licensed training corpus (this package aggregates varied sources with per-source licensing and some unknown upstream permissions). Also note that some generated targets are marked as human_verified:false, so downstream use should account for mixed verification status.
