I'm encountered an obvious recognition error when using best eng model

[Heermosi]

I'd like to upload the errornously recognized figures here
But it seems github cannot show it directly, so I've uploaded it to a repositoary.

The errornously recognized file can be viewed here:
https://github.com/Heermosi/misc/blob/master/0093-%3E0083.png

The eng best trained data given a 98% confidence to the 3rd digit over '8', obviously wrong.
For the other frames other than this pic, 9 seems to be recognized correctly again.

I think there might be some instability inside the model, shall you be looking into it?
And by the way I'm recognizing the printed characters.