How to improve the quality of tessdata by training lstm file using tesseract 4.1?

[yh l]

I've already trained the lstm files using Fine Tuning for a few characters, but result is not good. Some characters which I already add into .training_text file still can not recognize. Does any one have some opinions? THANKS.