# DocLayNet v1.2: human-annotated document layout segmentation

## Facts

- Provider: IBM Research, Deep Search team (Pfitzmann, Auer, Dolfi, Nassar, Staar)
- License: CDLA-Permissive-1.0 (https://github.com/DS4SD/DocLayNet/blob/main/LICENSE)
- Size: 87 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/docling-project/DocLayNet-v1.2
- Access: public, readable without an account
- DOI: https://doi.org/10.1145/3534678.3539043
- Category: Computer Vision
- Source: https://doi.org/10.1145/3534678.3539043
- Page: https://scigantic.com/archive/bf0cacc0-5473-5c5d-b563-604275206c8d
- Keywords: Computer Vision, image segmentation, benchmark, Parquet, Hugging Face

## About

DocLayNet is IBM Research's human-annotated benchmark for document layout analysis, published at KDD 2022. It gives page-by-page layout segmentation ground truth as bounding boxes and polygons for 11 classes (Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Section-header, Table, Text, Title) on 80,863 unique pages drawn from six document categories: financial reports, scientific articles, laws and regulations, government tenders, manuals and patents. Pages were labelled by trained annotators and a fraction are double- or triple-annotated, which lets users estimate annotation agreement as an upper bound on model accuracy. Version 1.2 extends the original release by embedding the source PDF of each page in a binary column alongside the page image, the bounding boxes, category ids, segmentation polygons, the PDF text cells that fall inside each box, and page and document metadata. Fixed train, validation and test splits avoid leakage of layout styles. Measured 2026-11-14: 85 Parquet shards (79 train-and-val, 6 test) in data/, about 39.8 GB in total. Open with the Hugging Face datasets library, pandas or pyarrow; it is the standard training set for layout detectors such as those used in Docling and for object detection models applied to PDFs. Licensed by IBM under the Community Data License Agreement Permissive 1.0; cite the KDD 2022 paper.

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("docling-project/DocLayNet-v1.2")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `data/test-00000-of-00006.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/bf0cacc0-5473-5c5d-b563-604275206c8d
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/bf0cacc0-5473-5c5d-b563-604275206c8d
- Over MCP: `get_schema_card` with id `bf0cacc0-5473-5c5d-b563-604275206c8d` at https://api.scigantic.com/mcp

## Related topics

- [Parquet datasets](https://scigantic.com/topics/parquet)
- [Benchmark datasets](https://scigantic.com/topics/benchmark)
- [Image segmentation datasets](https://scigantic.com/topics/image-segmentation)
- [Computer Vision datasets](https://scigantic.com/topics/computer-vision)
