Problem in training tesseract for vietnamese

[MinhQuan1909]

When I use tessseract to ocr the image above, I got some character error. Ex: Hòa (image) -> điòa (tesseract), Điều 1 (image) -> Điều L (tesseract).
So I am training tesseract with vietnamese to get better result. I create 30k images with https://github.com/Belval/TextRecognitionDataGenerator with 4 font time new roman (normal, bold, italic, bold italic). I trained with result BCER train=0.009%, BWER train=0.035%. But after I trained, some case got better, but some case got worse. Ex:
Better: điòa (tesseract before train) -> Hòa (after trained, correct)
Worse: 123/HD/NN (tesseract before train, correct) -> 123/HDANN (after trained)
Why some case got worse? Is my dataset small or something else?