Case Study

Why a model can look strong in development but weak in final evaluation

Case context: A company trains a system to classify customer support emails into topic categories. The model scores very well on the development set, so the team feels confident. After deployment testing, though, performance on the held-out evaluation set is much lower. On inspection, they discover that the development data came from one product line, while the final evaluation data came from another.

Question: If the development and evaluation sets were drawn from different distributions, why is it hard to pin down the cause of the low evaluation score? Name the main explanations the team should keep in mind.

Sample answer: When the two sets do not match, a low score on the evaluation set does not point to a single clear cause. The team cannot tell whether the problem comes from tuning too closely to the development set, from the evaluation set being intrinsically more difficult, or from the model already being near the best result it can achieve for this task.

Key points:

  • A mismatch between development and evaluation data makes the root cause unclear.
  • The model may have been tuned too specifically to the development set.
  • The evaluation set may be more challenging than the development set.
  • The model may be performing about as well as expected for the task.

Rubric: The answer should state that different data distributions make the diagnosis uncertain. It must identify the three explanations from the lesson: 1) over-tuning to the development set, 2) a harder evaluation set, and 3) the model already being near its attainable performance.

0

1

Updated 2026-08-12

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Machine Learning Yearning @ DeepLearning.AI

Related