User permissions required: ‘View Sources’ AND ‘View Labels’
Every time you train the platform on your data (i.e. labelling any verbatims), a new version of the model associated with your dataset is created. As these models are large and complex, previous versions are not automatically stored in our databases, purely as the storage requirements would be incredibly large.
The latest version of the model will always be readily available, but users are able to ‘pin’ a specific model version that they would like to save. They can also choose to 'tag' pinned models with a 'Live' or 'Staging' tag.
There are a couple of reasons for pinning a model version:
- Pinning a model gives you determinism over predictions, particularly for when you are using Triggers. This means that you can be confident of precision and recall scores for this version of the model, and future training events will not alter them (for better or worse)
- In the Validation page, users can see the validation scores for previous pinned model versions, allowing you to compare the scores over time and see how your training has improved your model
To pin a model version:
- Navigate to the models page via the left-hand dataset navigation bar
- Use the 'Save' toggle to save the current model version
Example model cards
To update the tag for a model version:
- Click the arrow next to 'Tags' on any pinned model
- Select 'Live' or 'Staging', depending on the status of the pinned model in any downstream deployments
Next: Deleting a pinned model