Question

A team iterating on a customer-support classification prompt has been tuning it repeatedly against the same 50-example test set. Their scores keep improving, but when they deploy, real-world accuracy is noticeably lower than their eval suggested. What is the most likely cause and best corrective action?

Accepted Answer

The prompt has been overfit to the small test set; they should maintain a separate held-out set that they do not tune against and use it to validate generalization.

Answer

The model temperature is too high; they should set it to 0 and the eval scores will match production.

Answer

The test set is too large to be reliable; they should shrink it to 10 examples for faster iteration.

Answer

Automated grading is inherently inaccurate; they should abandon the eval and rely on manual spot-checks instead.