Add an option to train a masked-language model
https://towardsdatascience.com/masked-language-modelling-with-bert-7d49793e5d2c
To fine-tune the pretrained model on a domain-specific data before the classification task.
Should be a separate application?