Datasets:

---

riotu-lab

/

SARD

like 13

Hello SARD team,

I am developing an Arabic OCR system for printed, historical, and low-quality scanned documents.

Could you please provide the official download links for the SARD dataset, including:

• training, validation, and test page images;
• the corresponding ground-truth/reference texts;
• image-to-text mapping files;
• annotations such as JSONL, XML, Markdown, ALTO, PAGE-XML, CSV, or metadata;
• the official train, validation, and test splits.

I understand that the dataset is listed under CC BY-NC-ND 4.0, so I would appreciate explicit written clarification on whether I may:

1. use SARD to train or fine-tune another OCR or vision-language model;
2. create internal augmented or degraded copies;
3. publish benchmark results and training/integration code;
4. publish LoRA adapters or trained/merged model weights;
5. use the resulting model in a future commercial product or paid API;
6. retain the original dataset privately without redistributing it.

Please also clarify:

• whether trained model weights are considered derivative works of the dataset;
• whether a separate commercial license or written authorization is available;
• the required attribution wording;
• whether any underlying source data has separate restrictions.

I previously contacted the team by email and am posting here in case this is the preferred communication channel.

Best regards,
Ahmad Sayed