Parseq recognition model is repeating characters

[hanshupe007]

After fine-tuning the parseq model is recognizing too many characters, in particular if the same character occurs multiple times in a row.
For example: O11001 is recognized as O111001, -942.333 as -942.3333, : 66719 as 666719 or UOOO as UOOOO (the words are clearly readable in standard Arial font). Another example:

73973 as 73973973

I trained with freeze-backbone on 25k synthetic data, and also tested (for now) on synthetic data. Did anyone experience similar issues?