The third phase, and the final step of the training process, is called ‘Refine’. The purpose of this stage is to refine your model and improve labels that are not performing as expected, as well as ensuring you have captured all of the labels and concepts that you are interested in.
Re:infer is designed to be completely transparent to users when it comes to model performance, and very flexible when it comes to improving performance in areas that require it. For any use case, you want to be confident that your model captures an accurate representation of what's in your dataset, and this phase of the training helps ensure that you can be.
This section of the Knowledge Base will cover in detail the steps outlined below, beginning with a detailed explanation of how Validation works, and how to understand the different aspects of model performance.
Validation - this step is about understanding and using Validation to see how your model is performing and how to improve it, with guidance from the Re:infer platform. This section includes detail about understanding and improving model performance.
'Check label' & 'Missed label' - these are two training modes (previously 'Teach' + reviewed filter) that help you improve label performance by reviewing previously reviewed verbatims where selected labels may have either been incorrectly applied or missed.
'Teach' (for unreviewed verbatims) - this is a step covered in more detail in the Explore phase (here), but it's another mode that can help improve label performance if needed by providing more varied training examples for a label. This mode shows you verbatims where the platform is unsure whether the selected label applies or not.
Check coverage - once you've correct any issues with label performance, you need to check that your model well covers you data and that a high proportion of all verbatims have good predictions. This step explains how to do that.
Next: How does validation work?