amrosama/al-lataif-al-musawwara-magazine-ocr

Al-Lataif Al-Musawwara Magazine OCR Pages

This dataset contains rendered Arabic magazine page images from Al-Lataif Al-Musawwara paired with page-level text and line-level bounding boxes. It is intended for OCR, document understanding, and VLM fine-tuning experiments.

Fine-Tuning Notebook

A standalone Google Colab notebook for DeepSeek-OCR 3B + TRL SFT is available here:

Open the fine-tuning notebook in Colab

The notebook can train on this dataset alone or on all three Arabic magazine OCR datasets.

Dataset Structure

Each row represents one page image.

Column	Description
image	Page image rendered as a Hugging Face image column.
image_id	Stable row id derived from the source path.
source_path	Original relative image path.
collection	Collection folder name.
volume	Source volume folder name.
page_number	Page number parsed from the filename.
width, height	Image dimensions in pixels.
ocr_text	Full page transcription text.
line_count	Number of detected text lines.
line_texts	Text from each detected line.
line_bboxes	JSON array of line polygon bounding boxes.
lines_json	JSON array containing line text and polygon bounding boxes.
instruction, response	Image-to-text fine-tuning pair. The instruction is Extract all text from this image.
task	Task label.

Splits

Split mode: volume.

Split	Rows	Parquet shards
train	3846	4
validation	32	1
test	32	1

Total rows: 3910

Usage

Because this dataset is large, streaming is the most convenient way to inspect or train from it:

import json

from datasets import load_dataset
from huggingface_hub import get_token

ds = load_dataset(
    "amrosama/al-lataif-al-musawwara-magazine-ocr"
)

sample = next(iter(ds["train"]))
image = sample["image"]
text = sample["ocr_text"]
line_texts = sample["line_texts"]
lines = json.loads(sample["lines_json"])
bboxes = json.loads(sample["line_bboxes"])

print(sample["image_id"], image.size, sample["line_count"])

For VLM fine-tuning, use image as the visual input, instruction as the prompt, and response as the target answer.

For non-streaming use on machines with enough disk/cache space:

from datasets import load_dataset
from huggingface_hub import get_token

ds = load_dataset("amrosama/al-lataif-al-musawwara-magazine-ocr", token=get_token())
sample = ds["train"][0]

Curation Notes

Images without matching valid OCR JSON were skipped. Common source/footer marks and angled watermark fragments were filtered from the text fields; the source page images are unchanged. The text and bounding boxes were generated automatically and may still contain mistakes from historical print quality, page skew, watermarks, stamps, low contrast, cropped text, or mixed page layouts.

Suggested Evaluation

Use held-out pages to evaluate full-page transcription quality, line detection coverage, and robustness across different publication years and scan conditions.

License & Attribution

This dataset is released under the CC BY 4.0 License.

In plain English:
You are completely free to use, share, adapt, or fine-tune models with this dataset for any purpose—academic or commercial. The only requirement is that you give credit by mentioning or citing this dataset.

Citation

If you use this dataset in your work or research, please cite it as:

@misc{al_lataif_ocr_dataset,
  author = {Amr Osama Abdellatif},
  title = {Al-Lataif Al-Musawwara Magazine OCR Dataset},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Repository},
  howpublished = {\url{[https://huggingface.co/datasets/amrosama/al-lataif-al-musawwara-magazine-ocr](https://huggingface.co/datasets/amrosama/al-lataif-al-musawwara-magazine-ocr)}}
}