Model doesn't support Arabic numbers or right-to-left rendering

#1
by saram1m - opened

While using the model with Arabic content, I noticed the following issues:

1.The model doesn't recognize Arabic-Indic digits (e.g., ูกูขูฃูค).

2.It doesn't handle right-to-left (RTL) layout properly, especially in structured outputs like tables.

This affects the quality and usability of extracted Arabic data in documents.


Please consider adding

1.Support for Arabic-Indic numeral normalization

2.Improved RTL text handling, especially in tables

Yes, I can confirm that these issues exist when using the DREX-062225-7B-exp model. While the model performs well overall with Arabic text, it still has some difficulties with Arabic numerals and RTL text direction.

I hope these aspects can be improved, as the model is impressive and delivers great results โ€” further enhancements would make it even more effective for Arabic content.

@xuilz1 @saram1m
Noted,
The fix will be available in the future on the same model page. I'll update you all here once it's dropped.

Sign up or log in to comment