ChunkrAI/chunkr-reading-order-bench-oss

Chunkr Reading Order Bench - Open Source Subset

Open-source subset of the Chunkr Reading Order benchmark dataset, containing 733 professionally annotated documents with detailed reading order annotations across diverse document layouts.

This dataset benchmarks reading order detection models on complex, real-world documents including financial reports, legal contracts, research papers, medical records, and more. Each document includes ground truth reading order sequences essential for accurate document understanding and text extraction.

---

Benchmark Comparison

We evaluated our model against traditional heuristics, machine learning approaches, and state-of-the-art Vision-Language Models (VLMs):

Model	Exact Match	Kendall's τ	Spearman's ρ	Position Acc	Inference (ms)
Chunkr Reading Order	88.1%	98.2%	98.5%	96.1%	44.6
openai/gpt-5	74.8%	85.0%	85.3%	82.5%	15035.2
google/gemini-2.5-pro	72.4%	82.1%	82.5%	79.3%	4678.5
x-ai/grok-4-fast	68.8%	91.4%	92.5%	83.2%	5661.1
Pairwise XGBoost*	67.7%	92.8%	94.6%	84.7%	16.7
Pairwise LightGBM*	67.4%	92.9%	94.9%	84.3%	33.6
anthropic/claude-sonnet-4.5	53.8%	84.5%	87.1%	73.1%	919.4
XY-Cut	51.7%	84.4%	85.5%	68.5%	1.7
TBLR (Baseline)	50.3%	87.5%	90.4%	75.2%	1.7
qwen/qwen3-vl-235b-a22b-instruct	49.1%	79.7%	80.9%	66.7%	4668.5
z-ai/glm-4.5v	46.8%	74.5%	75.4%	61.0%	8968.0
LayoutLMv3 Base	9.4%	12.0%	13.0%	23.9%	44.2

*Pairwise models trained on Chunkr dataset with optimized features

Dataset Overview

Document Categories

• Financial (14.3%), Legal (13.1%), Medical (10.8%), Government (10.0%)
• Technical (5.0%), Billing (5.2%), Consulting (4.6%), Research (3.5%)
• Education, Tax, Procurement, Historical, Supply Chain
• Real Estate, Patent, Construction, Newspaper, Magazine, Textbook

Layout Diversity

• 49.4% Single-column documents
• 12.8% Double-column documents
• 34.4% Three-column documents
• 3.4% Four+ column documents

---

📊 Reading Order Complexity

The dataset spans a wide range of layout complexities:

Layout Type	Count	Percentage	Description
Single Column	362	49.4%	Simple linear reading flow
Double Column	94	12.8%	Academic papers, newsletters
Multi-Column (3+)	277	37.8%	Newspapers, magazines, complex reports

Element Classes (16 types)

Class	ID	Count	Description
Text Block	1	2,866	Main body text regions
List Item	11	1,748	Bulleted or numbered list elements
Title	8	1,112	Section and document titles
Page Number	12	470	Page numbering elements
Header	7	429	Document headers
Footer	2	402	Document footers
Table	14	392	Tabular data structures
Picture	13	377	Images and photographs
Form Region	4	351	Form fields and structures
Graphical Item	6	250	Charts, diagrams, illustrations
Caption	0	231	Figure and table captions
Footnote	3	156	Footnote text
Line Number	10	145	Line numbering (code, legal docs)
Formula	5	57	Mathematical expressions
Legend	9	41	Chart and diagram legends
Unknown	15	240	Unclassified elements

---

Dataset Structure

chunkr-reading-order-bench-oss/
├── images/                    # 733 document images (PNG format)
├── _annotations.coco.json     # COCO format annotations with reading order
├── metadata.csv               # Image metadata (doc_category, dimensions, elements, columns)
└── analysis/                  # Statistical analysis and visualizations
    ├── dataset_overview.png
    ├── reading_order_visualization.png
    ├── column_distribution_by_category.png
    ├── reading_order_sample_*.png
    └── ...

---

Reading Order Annotations

The dataset provides professionally annotated reading order sequences for each document, capturing the natural flow in which humans would read the content. This is critical for:

• Accurate text extraction from complex layouts
• Document understanding and information retrieval
• Maintaining semantic coherence in multi-column documents
• Preserving logical flow across mixed layout elements

The arrows indicate the ground truth reading order progression, showing how the document should be read to maintain semantic coherence.

---

Dataset Statistics

Raw Layout Examples

---

Usage

Load with Hugging Face datasets

from datasets import load_dataset

dataset = load_dataset("ChunkrAI/chunkr-reading-order-bench-oss")

for item in dataset['train']:
    print(item['doc_category'], item['num_columns'], item['num_elements'])

With Supervision

import supervision as sv

dataset = sv.DetectionDataset.from_coco(
    images_directory_path="images",
    annotations_path="_annotations.coco.json"
)

print(f"Dataset size: {len(dataset)}")
print(f"Classes: {dataset.classes}")

# Access reading order from annotations
for image_path, image, annotations in dataset:
    # Process annotations with reading order information
    pass

---

Data Attribution and References

The following tables summarize datasets and sources referenced for sampling, benchmarking, or cross-domain balancing in Chunkr Reading Order Bench OSS. All sources are publicly accessible for research or under permissive licenses (MIT, CC-BY, public domain).

---

🧾 Financial & Business Documents

Dataset / Source	Domain	Access Location
DocLayNet (IBM)	Finance, Legal, Patents, Technical	Hugging Face – ds4sd/DocLayNet
IIIT-AR-13K	Financial Reports, Business	IIIT Hyderabad / GitHub
Kleister Charity / NDA	Financial & Legal Documents	GitHub – applicaai/kleister-nda
SEC Filings (EDGAR)	Corporate Finance	SEC EDGAR Bulk API
McKinsey, BCG, Deloitte Insights Reports	Consulting, Business Strategy	BCG Insights / McKinsey / Deloitte Portals
SynFinTabs	Synthetic Financial Tables	Hugging Face
RVL-CDIP	Business Records, Legal	Hugging Face – aharley/rvl_cdip

---

⚖️ Legal, Government & Regulatory

Dataset / Source	Domain	Access Location
FUNSD	Government Forms	FUNSD Dataset Page
NIST Tax Forms (1988)	Tax / Payroll	NIST Special DB 2
CourtListener RECAP Archive	Legal Filings, Case Law	Free Law Project
Cook County Contracts & Amendments	Legal / Government	Cook County Open Data Portal
GovDocs1 Million Corpus	Government / Multi-Domain	DigitalCorpora Repository

---

🏥 Medical & Healthcare

Dataset / Source	Domain	Access Location
Synthetic Medical Forms / OMR Scanned Medical Forms	Healthcare / Admin	Hugging Face Datasets
Synthea Synthetic Patient Records	Healthcare (EHR)	AWS Open Data Registry
Blue Cross Blue Shield Sample EOBs	Healthcare / Insurance	Public Sample PDFs
Handwritten Rx (Kaggle)	Prescriptions	Kaggle Dataset

---

🏢 Real Estate, Construction & Engineering

Dataset / Source	Domain	Access Location
CubiCasa5K	Floor Plans, Real Estate	GitHub – CubiCasa5k
CISOL / FloorPlanCAD	Construction / Supply Chain	Zenodo DOI 10.5281/zenodo.10829550
Machine Learning STRUCTural Floor Plan	Engineering / CAD	GitHub Repository
Lease Agreements / HUD Forms	Real Estate / Legal	HUD.gov Forms
Property Records / PhilaDox	Land Records	PhilaDox OpenData Portal

---

📚 Academic, Historical & Educational

Dataset / Source	Domain	Access Location
DocBank	Academic (ArXiv Preprints)	GitHub – doc-analysis/DocBank
PubLayNet	Scientific Journals	IBM DAX Project / Kaggle
U-DIADS-Bib / DIVA-HisDB / HJDataset	Historical Documents	ICDAR / UniFr / Harvard HisDoc Projects
Internet Archive "Magazine Rack" & Textbook Scans	Magazines, Textbooks	Internet Archive
Standardized Exam Papers	Education	College Board, ACT Archives

---

🧮 Technical, Manufacturing & Miscellaneous

Dataset / Source	Domain	Access Location
PRImA Layout Analysis	Technical Publications	PRImA Dataset
Safety Data Sheets (SDS) Online Repositories	Manufacturing / Chemical	SDSManager / Chemical Safety
Disassembly BOM Dataset	Manufacturing / Electronics	Figshare
Tobacco3482 / Tobacco800	Business Records	Kaggle Dataset
Customs Import Declarations (Synthetic)	Logistics / Customs	GitHub Dataset

---

License

MIT License
Free to use for research and commercial purposes.
Where applicable, individual third-party datasets retain their original licenses, all of which allow redistribution or benchmarking under fair-use or research terms.

---

Citation

If you use this dataset or benchmark, please cite:

@dataset{chunkr_reading_order_bench_oss_2025,
  title     = {Chunkr Reading Order Bench - Open Source Subset},
  author    = {ChunkrAI},
  year      = {2025},
  publisher = {Lumina AI Inc.},
  license   = {MIT},
  url       = {https://huggingface.co/datasets/ChunkrAI/chunkr-reading-order-bench-oss}
}