README.md

---
license: cc-by-4.0
task_categories:
  - image-to-text
language:
  - ar
  - fr
  - en
tags:
  - ocr
  - arabic
  - handwriting
  - invoices
  - document-understanding
size_categories:
  - 10K<n<100K
pretty_name: Alhazen-OCR Data
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: eval
        path: data/eval-*
---

# Alhazen-OCR Data

`alhazen-ocr` is the training dataset behind
[`context212/alhazen-ocr`](https://huggingface.co/context212/alhazen-ocr), an
Arabic-first OCR vision-language model. It combines **license-clean** Arabic
OCR sources — synthetic documents, institutional invoices, and handwritten
text — into a single normalized `image + text` format, with a held-out eval
split for CER/WER benchmarking.

Quick links:
- 🤗 Model: [`context212/alhazen-ocr`](https://huggingface.co/context212/alhazen-ocr)
- 🛠️ Code (data pipeline, training, eval): [`github.com/context212/atlas-ocr`](https://github.com/context212/atlas-ocr)
- 📊 External benchmark: [`ahmedheakl/arocrbench_khatt`](https://huggingface.co/datasets/ahmedheakl/arocrbench_khatt)

## What's in the dataset

Each row is one document image with its full transcription:

- `image`: the document image (page, crop, or line, depending on source)
- `text`: the transcription in natural reading order
- `source`: the upstream dataset the row came from

| Split | Rows |
|---|---|
| train | 52,726 |
| eval | 1,000 |

## Source composition

The mix targets 60% synthetic documents, 20% invoices, 20% handwriting. Where
a source has fewer rows than its target share, all available rows are used.

| Source | Content | Rows used | License |
|---|---|---|---|
| [`loay/arabic-ocr-synthetic-scans-faker-300k`](https://huggingface.co/datasets/loay/arabic-ocr-synthetic-scans-faker-300k) | Synthetic Arabic document scans | 45,000 | CC-BY-4.0 |
| [`KhalfounMehdi/arabic-latin-invoices-synthetic`](https://huggingface.co/datasets/KhalfounMehdi/arabic-latin-invoices-synthetic) | Invoices (ar/en/fr) | 4,014 | CC-BY-4.0 |
| [`johnlockejrr/KHATT_v1.0_dataset`](https://huggingface.co/datasets/johnlockejrr/KHATT_v1.0_dataset) | Handwritten Arabic (KFUPM) | 4,672 | MIT |
| [`sherif1313/Historical-Arabic-Handwritten-OCR`](https://huggingface.co/datasets/sherif1313/Historical-Arabic-Handwritten-OCR) | Historical handwritten pages | 40 | Apache-2.0 |

**Deliberately excluded** on license grounds (non-commercial):
`aamijar/muharaf-public` (CC-BY-NC-SA), `freococo/*` (CC-BY-NC-ND). Every row
in this dataset comes from a commercially usable source.

## Build pipeline

`scripts/build_dataset.py` in the repo above:

1. Downloads each source from the Hub (with retry/backoff for rate limits)
2. Normalizes every source to the `image + text` schema, renaming columns as
   needed and casting to a common feature type
3. Samples each source to its target share of the mix (seed 42, reproducible)
4. Shuffles, splits off 1,000 rows for eval, and pushes both splits

Rebuild with:

```bash
uv sync
HF_TOKEN=hf_... uv run python scripts/build_dataset.py
```

## Intended use

- Training / fine-tuning Arabic OCR vision-language models
- Evaluating OCR robustness across printed, invoice, and handwritten Arabic
- The `eval` split is the held-out set used in the model card of
  `context212/alhazen-ocr` — do not train on it

Not intended for:

- Languages other than Arabic/French/English (coverage is Arabic-first)
- High-stakes transcription without further validation

## Limitations

- Roughly 85% of rows are synthetic documents; real scanned and handwritten
  Arabic is a small minority, so models trained on this mix alone may
  underperform on real handwriting (this shows up in the model's external
  benchmark scores — see the model card)
- The historical handwriting source contributed only 40 usable image/text
  pairs at build time
- Texts are not re-verified against the images; upstream transcription errors
  carry through

## Licensing / Terms of Use

All sources are permissively licensed for commercial use (CC-BY-4.0, MIT,
Apache-2.0). The dataset as a whole is released under CC-BY-4.0. If you use
it, credit the upstream sources listed above.

## Citation

```bibtex
@misc{alhazen_ocr_2026,
  title  = {Alhazen-OCR: an open Arabic-first OCR vision-language model},
  author = {Context212},
  year   = {2026},
  url    = {https://huggingface.co/context212/alhazen-ocr}
}
```