metadata
license: mit
task_categories:
- image-segmentation
tags:
- arabic
- ocr
- document-understanding
- layout-analysis
- table-recognition
- chart-recognition
- vqa
- benchmark
KITAB-Bench is a comprehensive benchmark for Arabic OCR and document understanding. It evaluates performance across multiple tasks, including OCR, layout detection, table recognition, chart extraction, and visual question answering (VQA). The benchmark includes diverse document types from 9 major domains and 36 sub-domains, offering a robust evaluation of Arabic document processing capabilities.
Key Features:
- Diverse Datasets: Covers various document types (PDFs, handwritten text, tables, charts, etc.).
- Multiple Tasks: Evaluates OCR, layout detection, table recognition, chart extraction, and VQA.
- Comprehensive Evaluation: Uses established and novel metrics tailored to Arabic document processing.
- Strong Baselines: Provides performance results for leading models, including GPT-4o and Gemini.
Paper: KITAB-Bench: A Comprehensive Multi-Domain Benchmark for Arabic OCR and Document Understanding
Code: https://github.com/mbzuai-oryx/KITAB-Bench
Project Page: https://mbzuai-oryx.github.io/KITAB-Bench/