README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: apache-2.0
dataset_info:
  features:
    - name: image
      dtype: image
    - name: image_id
      dtype: string
    - name: style
      dtype: string
    - name: category
      dtype: string
    - name: source
      dtype: string
    - name: text
      list: string
    - name: word_count
      list: int32
    - name: total_words
      dtype: int32
    - name: bboxes
      list:
        list: int32
  splits:
    - name: train
      num_bytes: 402789351
      num_examples: 1272
  download_size: 399125040
  dataset_size: 402789351

DuwatBench: Bridging Language and Visual Heritage through an Arabic Calligraphy Benchmark for Multimodal Understanding [EACL 2026 (Main)🔥]

Shubham Patle ^1*   Sara Ghaboura ^1*   Hania Tariq ²   Mohammad Usman Khan ³  
Omkar Thawakar ¹   Rao M. Anwer ¹  Salman Khan ^1,4

¹Mohamed bin Zayed University of AI    ²NUCES    ³NUST    ⁴Australian National University

^{*Equal Contribution}

If you like our project, please like it ❤️ and give us a star ⭐ on GitHub.

Latest Updates

🔥🔥 [04 Jan 2026] 🔥🔥 DuwatBench accepted to EACL 2026 Main track.
🔥 [22 Jan 2026] DuwatBench, the Arabic Calligraphy Benchmark for Multimodal Understanding is released.
🤗 [23 Jan 2026] DuwatBench dataset available on HuggingFace.

---

✨ Overview

DuwatBench is a comprehensive benchmark for evaluating LMMs on Arabic calligraphy recognition. Arabic calligraphy represents one of the richest visual traditions of the Arabic language, blending linguistic meaning with artistic form. DuwatBench addresses the gap in evaluating how well modern AI systems can process stylized Arabic text.

Fig.1.a. Proportional breakdown of calligraphic styles in the DuwatBench dataset	Fig.1.b.Proportional breakdown of textual categories, covering religious and non-religious themes

---

🌟 Key Features

Key Features of TimeTravel

• 1,272 curated samples spanning 6 classical and modern calligraphic styles
• Over 9.5k word instances with approximately 1,475 unique words spanning religious and cultural domains
• Bounding box annotations for detection-level evaluation
• Full text transcriptions with style and theme labels
• Complex artistic backgrounds preserving real-world visual complexity

---

🏗️ DuwatBench Creation Pipeline

The DuwatBench dataset follows a structured pipeline to ensure the accuracy, completeness, and contextual richness by style and category.

Figure 2. End-to-end pipeline for constructing DuwatBench, from data collection and manual transcription with bounding boxes to multi-tier verification and style/theme aggregation.

---

✒️ Calligraphic Styles

Style	Arabic	Description
Thuluth	الثلث	Ornate script used in mosque decorations
Diwani	الديواني	Flowing Ottoman court script
Naskh	النسخ	Standard readable script
Kufic	الكوفي	Geometric angular early Arabic script
Ruq'ah	الرقعة	Modern everyday handwriting
Nasta'liq	النستعليق	Persian-influenced flowing script

---

🧐 DuwatBench Dataset Examples

Figure 2. End-to-end pipeline for constructing DuwatBench, from data collection and manual transcription with bounding boxes to multi-tier verification and style/theme aggregation.

---

🔍 Qualitative Evaluation and Results

Figure 3. Qualitative results comparing open- and closed-source models on DuwatBench calligraphy samples.

---

📚 Citation

If you use DuwatBench dataset in your research, please consider citing:

@misc{patle2026duwatbenchbridginglanguagevisual,
      title={DuwatBench: Bridging Language and Visual Heritage through an Arabic Calligraphy Benchmark for Multimodal Understanding}, 
      author={Shubham Patle and Sara Ghaboura and Hania Tariq and Mohammad Usman Khan and Omkar Thawakar and Rao Muhammad Anwer and Salman Khan},
      year={2026},
      eprint={2601.19898},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.19898}, 
}

---

⚖️ License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

The dataset images are sourced from public digital archives and community repositories under their respective licenses.

---

🙏 Acknowledgments

• Digital archives: Library of Congress, NYPL Digital Collections
• Community repositories: Calligraphy Qalam, Free Islamic Calligraphy, Pinterest
• Annotation tool: MakeSense.ai
• Arabic NLP tools: CAMeL Tools

---

📧 Contact

For questions or issues, please:

• Open an issue on GitHub
• Contact the authors at: {shubham.patle, sara.ghaboura, omkar.thawakar}@mbzuai.ac.ae

---