Essay

How to investigate a dataset mismatch after a dev-set drop

Question: A classification model scores well on its training data but noticeably worse on the development set. What should you do first to start diagnosing the mismatch, and what should you compare between the two datasets?

Sample answer: When a model performs much better on the training set than on the development set, the first diagnostic step is to examine how the data differ across the two sets. In particular, compare the properties and distributions of the training and development examples to identify what has changed.

Key points:

  • Recognize that a training-versus-development performance gap suggests a data mismatch.
  • State the first recommended diagnostic action: examine the data differences.
  • Focus on comparing the properties and distributions of the training set and the development set.

Rubric: The response must state that when a mismatch is detected because development-set performance is poor, the recommended first step is to investigate which data properties and distributions differ between the training and development sets.

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