Case context: A product team has been using a 150-example review set for model iteration. After each experiment, engineers inspect the errors on that same set and adjust features, thresholds, and data filters. Over time, the model’s score on the review set has risen sharply, but its score on a separate untouched evaluation set of several thousand examples has barely changed and is still clearly lower.

Question: What does this pattern suggest about the review set, and what should the team do next?

Sample answer: The pattern suggests that the review set has become overfit. Because the team repeatedly made decisions based on the same small set, the model and the development process were tuned to that set’s quirks instead of to the broader data distribution. The right response is to expand the review set by adding more representative examples, either by relabeling additional data from the untouched pool or by collecting new labeled examples. A larger set will make future error analysis more reliable.

Key points:

• A large gap between the reviewed set and the untouched set indicates overfitting to the reviewed set
• Repeated manual tuning on a small fixed set can cause this problem
• The untouched set provides a more objective check on generalization
• The remedy is to increase the size and representativeness of the review set

Rubric: Full credit requires identifying overfitting to the reviewed set from the performance gap and recommending that the team enlarge the review set with more representative data.