Kuis Siklus Hidup Model

Jawaban dan penjelasan muncul saat Anda mengikuti kuis.

1. Which sequence best captures the lifecycle described in the guide?
   • Define the problem, prepare data, train or select, evaluate, deploy, monitor and retire
   • Train once, publish the model file, and close the project
   • Deploy first, then define the purpose and evaluation criteria
   • Monitor only until training loss falls, then discard the release record
2. Before building a model, what decision does the guide recommend making?
   • Choose a model architecture before identifying the service owner
   • Commit to a production rollout before defining unsuitable conditions
   • Define the purpose, accountable owner, and whether a model is needed at all
   • Use the most recent training dataset as the project objective
3. Why should a release record identify data, preprocessing, prompts and evaluation materials as well as model weights?
   • Those components cannot affect behavior once weights are fixed
   • Changes outside the weights can alter behavior and must be traced to a version
   • The model file alone proves which inputs served every production request
   • Only prompts, never feature pipelines, need versioning
4. A new model version is ready for release. What deployment practice follows the guide?
   • Replace the old version for every client without a comparison period
   • Retain the old version but do not define when to restore it
   • Measure only offline training accuracy after switching all users
   • Roll out gradually where practical, compare versions and keep a rollback route
5. Which operational cases should be tested before a model-backed service is treated as ready?
   • Only successful requests with all dependencies available
   • Startup, cancellation, timeouts, dependency failure and client compatibility
   • Only the model's average score on a development set
   • Only whether the training checkpoint can be loaded locally
6. A demand model trained on dollar prices receives cent prices from a new service. What problem does this illustrate?
   • Training-serving skew caused by incompatible feature processing
   • Evidence that the model checkpoint changed during deployment
   • A valid currency conversion that cannot affect forecasts
   • A reason to omit input-range checks from the service
7. After deployment, what should the team monitor and assign an owner to review?
   • Only whether the endpoint returns an HTTP success code
   • Only whether training loss can be reduced by another run
   • Only the storage size of the latest checkpoint
   • System health, task outcomes, alerts and incident-response decisions
8. Why does deleting a retired model file alone not complete retirement?
   • It always causes the trained weights to reappear automatically
   • It erases the evaluation records needed for the next version
   • Retained input data, cached outputs or dependent services may still remain
   • The model must continue serving a small traffic slice indefinitely