PaddlePaddle

/

PP-OCRv6_medium_rec

like 21

Follow

PaddlePaddle 3.32k

Thank you for reaching out and sharing such detailed context. It's great to see the interest in bringing Arabic support onto the PP-OCRv6 architecture.

As you mentioned, the financial and KYC use cases for Arabic OCR are highly significant. While our core team is currently balancing multiple pipeline updates, we are very open to prioritizing this if we can gather enough robust training data.

To make an official v6 Arabic model a reality sooner, we highly welcome community contributions on the data front. If you have access to any standard datasets, dictionary resources, or text layout samples that could be shared, please let us know. Collaborative efforts like this are exactly what drive the PaddleOCR community forward.

Thank you for the suggestion, and looking forward to your reply!

Best regards,
PaddleOCR Team

Thanks for the quick and positive response!

Just to clarify the ask: we're not looking for a domain-specific model (banking / KYC / ID).
What the community really needs is a solid general-purpose PP-OCRv6 Arabic recognition model on the v6 architecture accurate on standard printed Arabic out of the box, and something we can then fine-tune on our own in-domain data (KYC, IDs, invoices, etc.).
A strong general base is what unlocks all the downstream use cases.
On the data front happy to help. I can share a set of Arabic datasets that together cover printed, synthetic, handwritten, and full-document text, which should be enough to train a strong base model and benchmark it properly:

• https://www.kaggle.com/datasets/humansintheloop/arabic-documents-ocr-dataset
• https://huggingface.co/datasets/mssqpi/Arabic-OCR-Dataset
• https://huggingface.co/datasets/freococo/synth_shamela_ocr_arabic_books
• https://huggingface.co/datasets/loay/arabic-ocr-synthetic-scans-faker-300k
• https://huggingface.co/datasets/freococo/ocr_arabic_books
• https://huggingface.co/datasets/medyas/arabic-ocr-printed-500k
• https://huggingface.co/datasets/mssqpi/Arabic-OCR-Dataset

I hope this Dataset help us.

Best regards.

Thanks for the feedback!

Your point about the importance of a strong general base model is absolutely spot on. We are very excited about the potential of this model, and we would love to invite you to partner with us to co-build this PP-OCRv6 Arabic model.

With your deep understanding of the language and these datasets, combined with our experience in training the PP-OCR architecture, we can create something highly valuable for the entire community.

To make this collaboration smooth and successful, we have a preliminary idea of how we can team up:

Training & Architecture (Our side): We will handle the core training, tuning, and iteration of the PP-OCRv6 model. We can also provide our standard OCR data format guidelines (annotation templates, etc.) to ensure smooth integration.

Data Organization (Your side): Since you know these datasets best, we’d love for you to help preprocess and organize the raw data into the specified format.

Evaluation & Benchmarking (Your side): Most importantly, we hope you can help design and curate a representative evaluation set (test set). An independent test set from your side will ensure the model’s out-of-the-box accuracy is objectively and accurately measured.

Please let us know if you’d be interested in collaborating on this. If this sounds good to you, we can share the data format specifications and discuss the details to get things rolling!