Eleven19/arabic-surya-training

Arabic Surya OCR Training Dataset

439 block-level Arabic book images (~531×783 px) with ground truth text for fine-tuning Surya OCR v1 recognition model.

Dataset Structure

• 439 samples extracted from printed Arabic book pages
• Each sample: rendered text block image + ground truth text
• Block-level crops preserve Arabic script details (diacritics, ligatures)

Usage

from huggingface_hub import hf_hub_download
import zipfile, json, os

DATA_DIR = '/content/training_data'
os.makedirs(DATA_DIR, exist_ok=True)

zip_path = hf_hub_download(
    repo_id='Eleven19/arabic-surya-training',
    repo_type='dataset',
    filename='surya_arabic_training.zip',
    local_dir=DATA_DIR,
    local_dir_use_symlinks=False,
)

with zipfile.ZipFile(zip_path, 'r') as zf:
    zf.extractall(DATA_DIR)

with open(os.path.join(DATA_DIR, 'metadata.jsonl'), encoding='utf-8') as f:
    metadata = [json.loads(line) for line in f]

Training

Use the Colab notebook ar_surya_finetune_v2.ipynb included in this dataset:

Load the notebook from this dataset and run on T4 GPU (~2-3 hrs).

Source

Extracted from printed Arabic book pages. Each image is a text block detected by Surya's layout analysis, manually validated for text quality.