Ciągłe szkolenie i automatyczny quiz przekwalifikowujący

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. Co odróżnia ciągłe szkolenie od automatycznego wdrożenia?
   • Training can create and evaluate a candidate without promoting it
   • Every training run must replace production
   • Deployment occurs before any evaluation
   • Continuous training avoids new data
2. Co powinno się stać, gdy alert dryfowania uruchomi ponowne szkolenie?
   • Publish the candidate immediately
   • Check data quality and schema before treating it as a model trigger
   • Assume every alert indicates real concept drift
   • Delete the previous model
3. Jakie korzyści operacyjne może zapewnić zaplanowany cykl przekwalifikowania?
   • It guarantees the candidate is better
   • It always uses the newest labels immediately
   • It gives a predictable cadence and compute budget
   • It eliminates data validation
4. Dlaczego kandydata należy porównywać z wdrożonym modelem?
   • To make data lineage unnecessary
   • To avoid choosing metrics
   • To guarantee subgroup performance
   • To detect regressions relative to the current production baseline
5. Co powinno blokować awans, jeśli krytyczna podgrupa ulegnie regresowi pomimo ogólnej poprawy?
   • An automatic release based only on the aggregate score
   • A predeclared subgroup gate or required review
   • A deletion of the subgroup from reporting
   • A replacement of the test set
6. Which lineage records help reproduce a trained candidate?
   • Only the GPU serial number
   • Only the model's public name
   • Data snapshot, source revision, dependencies and parameters
   • A screenshot of the final metric
7. Why retain the prior model after promoting a candidate?
   • It guarantees identical predictions
   • It makes the candidate more accurate
   • It removes the need for monitoring
   • It enables a rollback if live behavior degrades
8. What should a pipeline do if a required data or evaluation gate fails?
   • Stop promotion and retain diagnosable failure evidence
   • Publish the candidate but hide the report
   • Skip the failed check on every later run
   • Replace the baseline automatically