ner.batch-train seeing the actual results?


For all the results from running ner.batch-train as showing below:


Are we able to access the "wrong" results? I would like to see which label(s) generate the most wrong results.

We don't currently have support for that unfortunately, no. You should be able to create a custom recipe that does this easily though. The source for the ner.batch-train recipe is provided within Prodigy, or you can clone the prodigy-recipes repo and start from there: . This should make it easy to output the examples you want to see.

A nice addition to the ner.batch-train command would be an option to save the incorrect predictions to a dataset. We'll consider adding that.

Thanks for your answer, we will try to work together a solution. Another question we are facing: by running ner.batch-train, and leveraged another file as the evaluation source, we are seeing that the "correct" and "Right" and "Wrong" total of over 3000, but it only "using 731 for training"? Is there something wrong with my set up of the command? I double checked and am certain that I did not load the wrong dataset.