NER train curve with label stats?

Is it possible to combine train-curve and label-stats such that the output provides the train-curve plots for each label individually?

Also is there a way to save the metrics from train-curve in a file?

Hi! At the moment, the per-label stats are only available in the regular training, since it'd otherwise get very verbose very quickly. The per-label stats can also be a bit less representative when training with small portions of the data, because you can easily end up with very few instances of a given label.

That said, you can take a look at the implementation in recipes/ (you can run prodigy stats to find the location of your Prodigy installation) and make a small adjustment to how the training is run in train_curve. If you change the call to the _train helper like this, it'll show all detailed stats for each training run, including the per-label stats:

baseline, scores = _train(
    config, gpu_id=gpu_id, overrides=overrides, silent=False, show_label_stats=True

If you just want to save the output in the same format it's shown on the CLI, one option would be to just redirect the output to a file:

prodigy train-curve ... > train_curve.txt