Non-Latin language is reversed in recognition

[bely66]

I trained a recognition model in Persian. The issue is when the model predicts the output the text is reversed for some reason.

Any idea how can I resolve that?

for example: The right word is: احمد The predicted word is: دماح

Which is fine if it's consistently reversing it, the confusing thing is that it sometimes gets the text correctly so when I reverse the text i end up messing with the correct ones

Anyone know why this happens?