README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: question
      dtype: string
    - name: answer
      dtype: string
  splits:
    - name: train
      num_bytes: 1473584741
      num_examples: 200
  download_size: 1276133511
  dataset_size: 1473584741
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - table-question-answering
license: apache-2.0

KITAB-Bench: A Comprehensive Multi-Domain Benchmark for Arabic OCR and Document Understanding

KITAB-Bench is a comprehensive benchmark for evaluating Arabic OCR and document understanding capabilities. It covers nine major domains and 36 sub-domains across 8,809 samples, including diverse document types like PDFs, handwritten text, structured tables, and financial & legal reports. The benchmark evaluates performance across various tasks, including OCR, layout detection, table recognition, chart extraction, and PDF conversion. Novel evaluation metrics are introduced to ensure rigorous assessment across multiple dimensions.

Key Highlights:

• 9 major domains & 36 sub-domains across 8,809 samples.
• Diverse document types: PDFs, handwritten text, structured tables, financial & legal reports.
• Strong baselines: Benchmarked against Tesseract, GPT-4o, Gemini, Qwen, and more.
• Novel evaluation metrics: Markdown Recognition (MARS), Table Edit Distance (TEDS), Chart Data Extraction (SCRM).

Please see paper & code for more information:

• Paper
• Code
• Project Page