metadata
dataset_info:
features:
- name: image
dtype: image
- name: source
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 4555446
num_examples: 800
download_size: 4393010
dataset_size: 4555446
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: apache-2.0
task_categories:
- image-to-text
tags:
- arabic
- ocr
- document-understanding
- multi-modal
KITAB-Bench is a comprehensive benchmark for Arabic OCR and document understanding, evaluating performance across multiple domains and tasks, including text recognition, layout detection, table and chart extraction, and more. It includes 8,809 samples across 9 major domains and features high-quality human-labeled annotations. This dataset provides image data, source information, and corresponding text transcriptions.
Please see paper & code for more information: