One small difference is that we fixed the way the totals were reported in the meantime – so it's actually 372 examples (minus one example that was ignored). I'll adjust this accordingly in the docs example. Re-ran the experiments and the results look very similar with the latest versions of Prodigy and spaCy. I'm getting 74.598 (and 81.646 if I'm using the pretrained tok2vec weights with --init-tok2vec).
Btw, if you're looking for more end-to-end tutorials with data, check out the following projects:
Thank you very much for tolerating my laziness, for I just want to train the NER model beforehand without labeling them by myself.
With the labelled jsonl file, then I can train the model first, and if I am not content with the trained model, then I will label more.