---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: image
    dtype: image
  - name: texts
    sequence: string
  - name: bboxes_block
    sequence:
      sequence: float64
  - name: bboxes_line
    sequence:
      sequence: float64
  - name: categories
    sequence: int64
  - name: page_hash
    dtype: string
  - name: original_filename
    dtype: string
  - name: page_no
    dtype: int64
  - name: num_pages
    dtype: int64
  - name: image_width
    dtype: int64
  - name: image_height
    dtype: int64
  - name: collection
    dtype: string
  - name: doc_category
    dtype: string
  splits:
  - name: train
    num_bytes: 166529577.0
    num_examples: 400
  download_size: 162656888
  dataset_size: 166529577.0
  license: apache-2.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
task_categories:
- image-segmentation
---

**Please see paper & code for more information:**
- https://github.com/mbzuai-oryx/KITAB-Bench
- https://arxiv.org/abs/2502.14949

KITAB-Bench is a comprehensive multi-domain benchmark for Arabic OCR and document understanding.  It includes tasks such as text recognition (OCR), layout detection, line detection and recognition, table recognition, PDF-to-markdown conversion, chart-to-dataframe conversion, diagram-to-JSON conversion, and visual question answering (VQA).  The dataset contains a wide range of document types from various domains, with high-quality human-labeled annotations.  See the [project website](https://mbzuai-oryx.github.io/KITAB-Bench/) for more details.