KDoc-OCRBench Access Request
This dataset contains real-world Korean industrial documents and is not openly redistributable. Access is granted on a case-by-case basis for research and evaluation purposes only. Please fill in the information below to request access.
Log in or Sign Up to review the conditions and access this dataset content.
KDoc-OCRBench (Korean Document OCR Benchmark)
The first comprehensive Korean document OCR benchmark developed by ONTHEIT.
14,738 test cases across 804 Korean PDFs in 7 industrial document categories — designed to fill the gap in standardized Korean OCR evaluation.
Existing OCR benchmarks are primarily English-focused, making it difficult to evaluate model performance on Korean documents. KDoc-OCRBench addresses this gap with real-world Korean documents spanning contracts, medical records, financial forms, logistics paperwork, educational materials, government documents, and presentation slides.
Overview
- 804 Korean PDFs across 7 document categories
- 14,738 unit-test-style assertions covering text presence, text absence, table structure, and baseline quality
- Based on the olmOCR-bench methodology by Allen AI — each test is a focused assertion (e.g., "this text must appear", "this cell must be to the right of that cell") rather than a full-text comparison, making evaluation robust to formatting differences between models
- Korean-specific normalization for decorative spacing and multi-line table cells
Dataset Structure
KDoc-OCRBench/
├── pdfs/ # 804 PDF documents organized by category
│ ├── CorporateDocs/
│ ├── EducationalDocs/
│ ├── FinancialInsuranceDocs/
│ ├── LogisticsDocs/
│ ├── MedicalDocs/
│ ├── PresentationSlides/
│ └── PublicDocs/
├── long_tests.jsonl # Text presence/absence tests (10,137)
├── table_tests.jsonl # Table structure tests (4,147)
└── header_footer_tests.jsonl # Header/footer tests (454)
Test Types
| Type | Count | Description |
|---|---|---|
Text Presence (present) |
10,137 | Verifies specific Korean text appears in the output (with optional fuzzy matching) |
Text Absence (absent) |
454 | Verifies certain text (e.g., page headers/footers) is NOT in the output |
Table (table) |
4,147 | Validates table cell content and spatial relationships (up/down/left/right neighbors, column/row headings) |
| Baseline | 804 | Checks output is non-empty, not repeating, and contains valid characters (one per PDF) |
Document Categories
CorporateDocs (124 documents)
Corporate documents — employment contracts, service performance certificates, business plans, internal reports, meeting minutes, and more.
EducationalDocs (158 documents)
Educational documents — preventive education materials, student residence surveys, curriculum guides, school support procedures, admission assignment notices, academic transcripts, and more.
FinancialInsuranceDocs (64 documents)
Financial and insurance documents — securities issuance terms, payment notices, income deduction statements, credit transaction agreements, automatic transfer applications, insurance claim forms, and more.
LogisticsDocs (72 documents)
Logistics documents — simplified customs declarations, price declarations, import customs clearance guides, goods requisitions, transaction statements, multimodal bills of lading, and more.
MedicalDocs (118 documents)
Medical documents — medical records, medical opinions, diagnostic certificates, prescriptions, health checkup result notices, symptom and treatment records, medical fee receipts, and more.
PresentationSlides (99 documents)
Presentation materials — major initiative progress and plans, social investment materials, social impact reports, collaboration strategy decks, and more.
PublicDocs (169 documents)
Public documents — initial salary grade determination procedures, employment support program guides, performance indicator and weighting specifications, name change applications, leak detection reports, and more.
Benchmark Results
See the leaderboard and evaluation code at ONTHEIT-AI/BizOnAI-OCR.
Credits
- Korean document data collected and labeled by ONTHEIT
- Evaluation methodology inspired by olmOCR-bench by Allen AI
License & Access
This dataset contains real-world Korean industrial documents and is not openly redistributable. Access is granted on a case-by-case basis for research and evaluation purposes.
To request access, contact ONTHEIT at bizonai@ontheit.com. Granted users may use the dataset only for OCR evaluation and may not redistribute it.
- Downloads last month
- 17