metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: data
      dtype: string
    - name: uid
      dtype: int64
  splits:
    - name: train
      num_bytes: 99059837
      num_examples: 378
  download_size: 94572845
  dataset_size: 99059837
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - image-to-text
license: apache-2.0
tags:
  - arabic
  - ocr
  - document-understanding
  - multi-domain
  - benchmark

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. It covers 9 key tasks across 9 major domains and 36 sub-domains, including text recognition (OCR), layout detection, table recognition, PDF-to-Markdown conversion, chart-to-DataFrame extraction, diagram-to-JSON extraction, and Visual Question Answering (VQA). The dataset includes over 8,809 samples with high-quality human-labeled annotations. Novel evaluation metrics are introduced to accurately assess performance across these diverse tasks.

Key Features:

9 major domains & 36 sub-domains across 8,809 samples.
Diverse document types: PDFs, handwritten text, structured tables, financial & legal reports.
Evaluation across OCR, layout detection, table recognition, chart extraction, & PDF conversion.
Novel evaluation metrics: MARS, TEDS, SCRM, CODM.

For more details, including the complete methodology, evaluation metrics, and results, please refer to:

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/