HJDataset (mirror)

Mirror of the HJDataset (Japanese historical documents, Johns Hopkins University) layout analysis dataset used for DocLayout-YOLO downstream training experiments. Historical Japanese documents feature vertical text and right-to-left reading order.

Contents

• 8,000 files: document page images + YOLO-format .txt labels (11 layout classes).
• Where a source directory exceeds the HuggingFace limit of 10,000 files per directory, overflow files are sharded into shard_NN/ subdirectories (9,000 files each).

Exclusions

• labels.cache (Ultralytics label cache) intentionally excluded; .ipynb_checkpoints/ excluded.

Known data anomaly (recorded, not repaired)

One label file (990077217110203941_v17pt2_0659.txt) contains out-of-bounds normalized coordinates. It is uploaded as-is from the source; no labels were modified.

Provenance

• Used as one of four downstream benchmarks (D4LA / M6Doc / HJDataset / PRIMA-LAD) in the DocLayout-YOLO release training-pipeline verification (2026-09). Uploaded as-is; no source files modified.