Accuracy seems cannot be improved on my own dataset for im2text

Support
1

acc can only achieve about 70… when i do im2text training. dataset is generated from some latex (latext2png)
would some tell me how to handle this please?