Najwyraźniej sztuczna inteligencja do quizu dotyczącego monitorowania modeli typu open source

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. What does a reference-versus-current drift report compare?
   • Selected data distributions across two periods or datasets
   • Model accuracy without target labels
   • Only the training loss curve
   • The deployment and source-code branches
2. Why should a report record its reference window and feature selection?
   • They determine the model's hidden layers
   • Those choices affect which changes are measured and how results are interpreted
   • They ensure every drift score is a p-value
   • They eliminate sample-size effects
3. What can an unlabeled drift report establish?
   • That the model is unfair
   • That prediction accuracy has declined
   • A change in selected data distributions under configured methods
   • That the new data are incorrect
4. Which expectation does a data contract test verify?
   • Whether labels arrive immediately
   • Whether every model output is correct
   • Whether drift is always harmful
   • Whether inputs satisfy specified schema or quality expectations
5. Why can a strict drift threshold create noisy alerts?
   • The library ignores all features
   • Expected seasonality or sampling variation can trigger a change signal
   • Drift reports include ground truth automatically
   • Reference data are never needed
6. How should delayed target labels be used in monitoring?
   • Use unlabeled inputs as a substitute for outcome truth
   • Treat missing recent labels as incorrect predictions
   • Join them to predictions with matching model version and outcome window
   • Drop all historical predictions
7. Why version the report configuration and library version?
   • It changes the reference distribution
   • It guarantees the same input data forever
   • It removes the need for labels
   • Defaults and methods can change, affecting reproducibility
8. How should a team respond to a failed data-quality test?
   • Investigate whether the input violates a meaningful contract or reflects an expected change
   • Automatically retrain on the data
   • Ignore the test for every future run
   • Declare the model inaccurate without labels