Improve dataset card and metadata

This PR improves the dataset card by adding a more detailed description of KITAB-Bench, including its key features, tasks, and domains. It also adds metadata for the task category, language, and license, improving the dataset's discoverability and usability on the Hub.

README.md

@@ -13,13 +13,53 @@ dataset_info:
     num_examples: 500
   download_size: 82620782
   dataset_size: 97795630.0
+  language:
+    - ar
+  license: mit
 configs:
 - config_name: default
   data_files:
   - split: train
     path: data/train-*
+task_categories:
+- table-question-answering
+- visual-question-answering
+- image-to-text
+- document-question-answering
+- charts-visual-question-answering
+- diagram-visual-question-answering
+tags:
+- ocr
+- arabic
 ---
 
-**Please see paper & code for more information:**
-- https://github.com/mbzuai-oryx/KITAB-Bench
-- https://arxiv.org/abs/2502.14949
+# <img src="static/images/kitb.png" width="50" > KITAB-Bench: A Comprehensive Multi-Domain Benchmark for Arabic OCR and Document Understanding
+
+KITAB-Bench is a comprehensive multi-domain benchmark designed to evaluate Arabic Optical Character Recognition (OCR) and document understanding capabilities. It addresses the unique challenges posed by the Arabic language, including its cursive script, right-to-left text flow, and complex typographic variations.  The benchmark covers a wide range of document types, from PDFs and handwritten text to structured tables and financial reports, and assesses performance across various tasks, including OCR, layout detection, table recognition, chart extraction, and visual question answering (VQA).
+
+**Key Features:**
+
+* **9 Major Domains & 36 Sub-domains:** Covering diverse document types and scenarios.
+* **8,809 Samples:** Providing a substantial dataset for robust evaluation.
+* **High-Quality Annotations:** Ensuring fair and accurate assessment.
+* **Novel Evaluation Metrics:** Introducing metrics like MARS (Markdown Recognition Score), TEDS (Table Edit Distance Score), and SCRM (Chart Representation Metric) for comprehensive evaluation.
+* **Strong Baselines:** Benchmarked against leading models such as Tesseract, GPT-4o, Gemini, Qwen, and more.
+
+**Benchmark Tasks:**
+
+KITAB-Bench evaluates performance on the following key tasks:
+
+1. **Text Recognition (OCR):** Printed & handwritten Arabic OCR.
+2. **Layout Detection:** Identifying text blocks, tables, figures, and other elements.
+3. **Line Detection & Recognition:**  Detecting and recognizing individual text lines.
+4. **Table Recognition:** Extracting information from structured tables.
+5. **PDF-to-Markdown:** Converting Arabic PDFs to structured Markdown format.
+6. **Charts-to-DataFrame:** Extracting data from various chart types.
+7. **Diagram-to-JSON:** Extracting information from diagrams like flowcharts and Venn diagrams.
+8. **Visual Question Answering (VQA):** Answering questions about Arabic documents.
+
+**Please see the paper & code for more information:**
+
+* [Paper on arXiv](https://arxiv.org/abs/2502.14949)
+* [Code on GitHub](https://github.com/mbzuai-oryx/KITAB-Bench)
+* [Project Page](https://mbzuai-oryx.github.io/KITAB-Bench/)