I have found that a useful trick when training a model is to keep adjusting the subtitle sequence.

This way, you can get higher accuracy at the prediction stage.