Google

A bird-identification app is evaluated on a dev/test set containing mostly clear daytime photos of robins and finches, while users mainly upload dim phone photos of backyard birds at dusk. Because the evaluation and real-user inputs come from different distributions, dev/test performance may give a misleading picture of deployment performance.

Bird-Identification App Dev/Test Distribution Mismatch

Recognizing a Dev/Test Distribution Shift

A Dev/Test Set Should Match _____ Data

Distribution Shift in a Wildlife Classifier

Finding a Dataset Distribution Mismatch

Question:
A photo app is evaluated using a dev/test set made mostly of full-grown dogs, but after launch many users upload puppy photos. Explain why the original dev/test set may be misleading.

Sample answer:
The dev/test set was built around one kind of input: images of adult dogs. That makes it a reasonable check for performance on that population, but it does not match what the app actually receives after deployment if puppy photos become common. Because the real input distribution is different, performance on the original dev/test set may not predict real-world behavior. In that situation, the dev/test set is not representative of the target distribution.

Key points:
- The dev/test data mostly contains adult dog images
- Real users upload many puppy images
- The test distribution differs from actual use
- A mismatch makes the dev/test set unrepresentative

Rubric:
The response must explain that a dev/test set built around adult dogs does not reflect real usage if the deployed app receives many puppy images, and it must clearly conclude that the dev/test distribution is not representative of the actual data distribution.

Why a Test Set Can Fail to Reflect Real Usage

Case context:
You manage a plant-identification app used by gardeners. The original dev/test set mostly contained clear photos taken in bright daylight. After a major product launch, the logs show that many users are uploading blurry images taken indoors under artificial light.

Question:
What is the most likely problem with the current evaluation setup?

Sample answer:
The dev/test set is no longer representative of the images the system now receives in practice. Because the evaluation data still mostly reflects bright, clear outdoor photos, it may overstate how well the model performs on the indoor, blurry images that are now common in real use.

Key points:
- Original dev/test set mostly bright daylight photos
- Real usage now includes many blurry indoor photos
- Evaluation data no longer matches the operational data distribution

Rubric:
Correctly identifies that the dev/test set is not representative of the current user input distribution.

The Shift in Field Photos

Question:
A fraud model is checked using transactions from large in-store purchases, but after deployment most flagged cases come from small mobile payments. What does that change imply about the validation set?

Sample answer:
It means the validation and test data do not reflect the situations the model must handle after launch, so their evaluation results can be overly optimistic or otherwise misleading.

Key points:
- Validation/test data does not match the deployment setting
- Reported performance may not reflect real-world use

Rubric:
Must state that the validation/test data does not match the target production distribution.

Validation Data and Production Shift

A team builds a classifier for factory inspection photos. The dev and test sets are drawn from a fixed overhead camera inside a brightly lit packing line. After launch, however, most new inputs come from handheld phone photos uploaded by field inspectors in warehouses with uneven lighting. What does this tell you about the dev/test set?

Learn Before

Related