[Help] After training PP-OCRv5_server_rec with Vietnamese data, it no longer recognizes Japanese text

[tttiuem2k3]

Hello PaddleOCR team,

I need some help with an issue I encountered during training.
When I train the PP-OCRv5_server_rec model with Vietnamese data, it performs very well for Vietnamese recognition.
However, after training, the model forgets Japanese — it fails to recognize Japanese text correctly (characters are either misrecognized or displayed as squares).

Here’s what I did:

Kept the original PP-OCRv5_server_rec architecture

Used a custom Vietnamese dictionary (vi_dict.txt) during training

Confirmed that the Japanese dictionary still exists and is complete

Despite that, Japanese OCR no longer works properly, while the original pretrained model could handle multiple languages (including Japanese) without any problem.

👉 My question is:
How can I train the model with Vietnamese data without losing the multilingual capability (especially for Japanese)?

Thank you very much for your help!

[liuhongen1234567]

Hello, it is recommended to use PP-OCRv5_server_rec to scrape some Japanese data and add it to the training set for mixed training.