OmniDocBench: 1,355 Annotated PDF Pages for Diverse Document Parsing

OmniDocBench contains 1,355 PDF pages across 9 document types and 3 languages, developed by OpenDataLab for evaluating document parsing systems. The dataset provides 15 block-level layout annotations for diverse materials including academic papers, financial reports, and handwritten notes.

Use Cases

• Evaluating layout analysis performance using 15 block-level location annotations
• Benchmarking OCR systems across 3 language types including English and Chinese
• Testing parsing accuracy on complex layouts found in newspapers and textbooks

Strengths

• 1,355 annotated PDF pages
• 15 block-level layout categories for precise evaluation
• Includes diverse document types such as handwritten notes and financial reports

Limitations

• Small sample size of 1,355 records optimized for benchmarking rather than training
• Limited to 3 language types

Provenance

Source

OpenDataLab (Arxiv: 2412.07626)

Collection Method

Annotated

Freshness

Last updated September 2025.