The model rollback feature allows us to revert back to a previous version of our model, allowing us to reset the training data (for both label and entity annotations) to the annotations used to train this model version.
It is important to note that we can only roll back to pinned versions of models.
How to use this feature
On the 'Models' page, the model rollback icon will be available on all pinned versions of our model. To proceed with the model rollback, click the rollback icon on the model version you want to revert back to.
An image of the model rollback button on the 'Models' page
It is important to note that the current trained model version will automatically be pinned as a backup but any annotations captured by a model version that is currently still training will be lost.
We recommend allowing the current model version to finish training before rolling your model back. A popup module will come up to remind us of this, after clicking the rollback button. If we would like to proceed, we can click the 'Reset' button.
An image of the popup asking if we would still like to proceed with the rollback
If the model rollback has successfully kicked off, a banner will appear in the bottom right corner letting us know that the process has kicked off.
A banner indicating that the rollback has kicked off
While the model is rolling back, we will not be able to modify the dataset. This means that we will not be able to train our model during this time, and apply any labels or entities to verbatims. A warning indicator will show up at the top, letting us know that the model is currently being rolled back.
A popup indicating that the model is currently rolling back and modifications can't be made
If we try to modify our dataset, the following banner will appear in the bottom right corner of our screen, and any verbatims we try to annotate will not have the label or entity applied to it until the model rollback has complete.
A banner indicating that we have tried to modify a dataset, and we cannot do so during the model rollback
Although the rollback feature is here to help us roll back to a previous version of a model if we've made any major mistakes in our model training, we should not rely too heavily on it.
Instead, we should be ensuring that we are following the proper model training methodology correctly the first time, as this will save us time in the long-run.
Previous: Maintaining a model in production