Fine-Tune TrOCR on Arabic

I am trying to fine-tune TrOCR on Arabic, I followed TrOCR Fine-Tune Tutorials by @NielsRogge, choose a suitable encoder/decoder option suggested by @NielsRogge on discussion board but I get very high WER score and the generated text is gibberish, I tried the pytorch implementation and it generated text prediction to continue the sentence instead of writing the sentence in the image.

Notebook of Code

It seems you have a configuration issue

how should
I fix it?