---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: text
    dtype: string
  - name: source
    dtype: string
  splits:
  - name: train
    num_bytes: 13421862.0
    num_examples: 200
  download_size: 13211013
  dataset_size: 13421862.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: apache-2.0
task_categories:
- image-to-text
---

KITAB-Bench is a comprehensive multi-domain benchmark for Arabic OCR and document understanding.  It evaluates the performance of traditional OCR, vision-language models (VLMs), and specialized AI systems on diverse document types including PDFs, handwritten text, structured tables, financial & legal reports, and more.  The benchmark includes nine major domains across 8,809 samples and offers novel evaluation metrics such as Markdown Recognition Score (MARS), Table Edit Distance Score (TEDS), and Chart Representation Metric (SCRM).

**Key Features:**

* 9 major domains & 36 sub-domains
* 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: MARS, TEDS, SCRM, and more

**Please see paper & code for more information:**
- [Paper](https://arxiv.org/abs/2502.14949)
- [Project Page](https://mbzuai-oryx.github.io/KITAB-Bench/)
- [Github Repository](https://github.com/mbzuai-oryx/KITAB-Bench)