Learn how and when to activate “not for training” to avoid adding false training data to our machine-learning models.
How does it work?
Every document validated by a user is automatically added to the training data, which the machine learns from. Wrong validations lead to false training data, which in turn reduces extraction quality. To correct incorrect training data, you can either reset and re-validate the document or remove it from the training data. Documents can be removed before, during, or after validation.
Before validation
Documents can be uploaded with the "not for training" flag during the upload process. For more information, see the related article.
During Validation
If a document needs to be removed from training during validation, users can click on the "Tags & Flags" tab and select the checkbox to remove the document from training.
List view
Documents that have already been completed can also be removed from training. This can be done by going into the document’s details view and removing the training flag in the "Tags and Flags" tab.
Whenever you’re about to finish the validation of a document, all fields should be validated properly according to their requirements. If you’re ever unsure whether you’ve fulfilled the requirements of a field, or if you’re working on an integration and decide to confirm the extraction results without fully checking them (to save time), you can activate the “not for training” flag.
Once this flag is set, no training export will occur, and no false training data will be delivered to our machine-learning models.