Saving annotations when the --update
option using ner.teach
results in the following error:
if span["answer"] == "accept" KeyError: 'answer'
.
Without the --update
option, saving annotations works fine.
Saving annotations when the --update
option using ner.teach
results in the following error:
if span["answer"] == "accept" KeyError: 'answer'
.
Without the --update
option, saving annotations works fine.
Thanks, it looks like this is a mistake – I'll fix this for the next release. In the meantime, you could just edit the recipe and remove that condition. In the case of ner.correct
, this doesn't matter anyway, because all spans are considered correct.
I noticed that train
no longer accepts the --binary
flag. Does that mean if I pass a dataset from ner.teach
it knows how to use the dataset? Or should I just update the model with the --update
flag?
is there any update on this? For me it's also not clear if this training mode is available on spacy 3 - or how to invoke it, using the config-file based approach. Would also be good to know if it is not planned to integrate it in the spacy3-based new version of prodigy. Thanks!
The new version of Prodigy will unify the training workflow so you won't have to specify --binary
anymore. However, we still need a small update in spaCy to properly allow updating the components with the respective contstraints defined in the binary feedback, so the results are currently expected to be worse in the nightly. Also see my comment here:
Edit: Just released a new nightly with better support from updating from binary annotations and a fix for the bug mentioned above.