Essay

Why mismatched evaluation data can mislead a project

Question: Discuss why a model team can be misled if its development and test data come from a different source than the data the model will face in production, and state what they should do.

Sample answer: If the holdout data comes from a distribution unlike the production data, the score on that holdout set may not predict real-world behavior. The team may optimize for the wrong pattern and believe the system is improving when deployment performance is not. The fix is to rebuild or revise the development and test sets so they better match the target distribution the model must handle.

Key points:

  • Holdout data should reflect the data the system will encounter.
  • A mismatch makes evaluation results unreliable for deployment decisions.
  • The correct response is to update the development and test sets to match the target distribution more closely.

Rubric: The response should explain that a mismatch between holdout and production data gives misleading evaluation signals because the holdout is not representative, and it should state that the development and test sets need to be revised to better match the target distribution.

0

1

Updated 2026-08-12

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Machine Learning Strategy

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Yearning @ DeepLearning.AI