M6Doc (private->public mirror)

Mirror of the M6Doc (Multi-form, Multi-type, Multi-layout, Multi-language, Multi-annotation, Multi-task) document layout analysis benchmark used for DocLayout-YOLO downstream training experiments.

Contents

• images/: ~19,000 document page images (.jpg + matching .npy descriptors). The first 10,000 files are stored flat; overflow files are sharded into images/shard_10/, images/shard_11/, ... (9,000 files each) because HuggingFace repos allow at most 10,000 files per directory.
• labels/: YOLO-format .txt label files (74 layout classes).

Exclusions

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

Data note

On 2026-09-26, 4,050 truncated JPEG files (missing EOI marker) in the source copy were repaired in place by re-encoding with ImageOps.exif_transpose(...).save(f, "JPEG", subsampling=0, quality=100) — byte-identical to the data-loading behavior of the original DocLayout-YOLO training pipeline (which performs the same repair on first scan). Before/after SHA256 for every repaired file is recorded in the training-verification manifest (m6doc_inplace_repair_manifest.json). No other source bytes 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).