Skip to main content
Version: 1.40

Performance evaluation with Actuals

Evaluate the performance of a Deployment by comparing predicted outcomes to what we call actuals; ground truth or real-world observations. This may help you determine a Deployment’s effectiveness and identify areas for improvement.

Collect actuals

We assume you have already created a Deployment and made predictions. Use the Deeploy API or Python client to add actuals.

To comply with KServe Data Plane Formatting, the actual format must match the following structure:

{ 'outputs': [<value>] }

or

{ 'predictions': [<value>] }

Monitor actuals

Monitor actuals for a Deployment on the Performance tab on your Deployment’s Monitoring page. Different metrics are available for classification, regression, and text generation models. Read Monitoring a Deployment for more details.

View actuals

View the actual for a specific prediction by clicking on a prediction on the Predictions page within a Deployment. Scroll down to view:

  • the token used to supply the actual
  • the observed value