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
Tags
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
Make the Dev Set Match the Main Improvement Goal
A Shared-Source Validation/Test Gap Usually Means Validation Overfitting
Why Different Dev and Test Distributions Complicate Diagnosis
Mismatched Validation and Test Splits Can Make Chance Matter More
If a model is tuned on a development set and then performs worse on a separate test set, even though both sets come from the same source distribution, what is the most likely explanation?
True or False: If validation and test data are drawn from different populations, a score gap between them always has one clear cause.
If a model has started fitting the validation set too closely and the training and validation data come from the same distribution, the usual remedy is to get more _____ data.
Why should the development set match the main goal of the project?
When validation and test data come from the same distribution, a strong validation score followed by a much weaker test score suggests overfitting to the validation set.
If the development set is being overused and the training and development data come from the same distribution, the practical fix is to collect more _____ data.
Match each development-and-test-set situation with its consequence for debugging.
Order the troubleshooting steps when validation performance is strong but holdout performance is weak.
Why a Test Set Can Score Lower Than a Dev Set
A test score from a different data distribution tells you exactly why the model failed.
After the development and test sets are set, the team will spend its effort improving _____ set performance.
Match each concept about dev and test set distributions to the correct description.
How to Choose Dev and Test Sets for Reliable Iteration
Explain what poor final performance means when the tuning set and final set do or do not match.
Why a model can look strong in development but weak in final evaluation
What is the diagnosis and remedy when test results lag far behind dev results under the same data distribution?