Record training results


Is there a way I can see the training results from the model folder. Like we get the output when we train the model on the command line. Is that information stored somewhere in the metadata of the model?

Not currently – but this is a good idea! The teach recipes could just add an "accuracy" to the model’s meta, similar to how spacy train does it.

In the meantime, you can easily add this yourself by writing to the model’s nlp.meta in the ner.batch-train recipe, just before saving out the model. For example:

if output_model is not None:
    model.nlp.meta['accuracy'] = {'ner': best_acc}
    # ...

When the model is saved out, the contents of nlp.meta will be saved in the meta.json. You could also use this to add other information to it – like the Prodigy dataset it was trained on etc.

If you just want to be able to see the accuracy without re-training, I posted a recipe here that evaluates the model against an arbitrary dataset without training

1 Like