Prediction on raw data is very similar as predicting on full datasets, the only change is that you add --raw_text as a paramenter. The expected input format whitespace separated word, with a newline to separate the sentences:
I choose you Machamp !
Machamp , use mega punch !
Assuming this text is saved in battle.txt, the command would be:
python3 predict.py logs/ewt/<DATE>/model.tar.gz battle.txt battle.pred.conllu --raw_text
Note that this is only works for sequence labeling tasks at the moment, and for now is its also safer to use it with --dataset. If you use other task types we suggest to just add underlines (_) as labels.