Test Data
What is test data?
Test data is a collection of inputs and (optionally) expected outputs used to evaluate the performance and quality of an AI system. In AI development, test data helps validate that models and prompts work correctly across a representative range of scenarios.
Test data should be separate from development data to prevent overfitting and ensure unbiased evaluation. This separation ensures that improvements observed during development actually generalize to new, unseen examples.
How is test data different from a dev set?
Teams typically maintain two types of evaluation datasets:
Dev set (development set): A smaller collection used during active development for rapid iteration and experimentation. Teams run evals frequently on the dev set while tuning prompts, adjusting models, or testing new approaches.
Test set (test data): A larger, more comprehensive collection reserved for final validation before release. The test set confirms that improvements seen in the dev set hold across a broader range of examples and edge cases.
This separation prevents teams from inadvertently optimizing their system for specific examples in the dev set, which could lead to overfitting.
Why must test data be protected?
Test data integrity is critical for valid evaluation. If test data becomes contaminated—either by accidentally including it in training data or by repeatedly adjusting the system based on test results—the evaluation becomes unreliable.
Data leakage occurs when information from the test set influences development decisions, undermining the test set's ability to provide an unbiased assessment of system performance.
Teams should treat test data as a holdout set, using it only for final validation before significant releases rather than for iterative development decisions.
Learn more:
- How I Designed & Implemented Evals for Product Talk's Interview Coach
- Q&A Session: Building AI Evals for the Interview Coach
- Debugging AI Products: From Data Leakage to Evals with Hamel Husain
Related terms: