For users who would like to retrain the network on a different corpus, there is mo documentation; it would be neat to have documentation on the corpus format and training pipeline.