RAIDP01-BP01 Identify evaluation datasets needed to measure system performance against release criteria
Work backwards from your release criteria to identify the specific evaluation datasets needed to test each one. Validate that each dataset has the right characteristics for its purpose (for example, demographic labels for fairness testing, harmful content examples for safety testing, and sufficient sample sizes for statistical confidence). Track mappings between datasets and criteria so you can verify complete coverage and maintain traceability between your release criteria and testing approach.
Level of risk exposed if this best practice is not established: High
Implementation considerations
-
For each release criterion, develop a dataset design that specifies the required data sources and data labels. Use your analysis of intrinsic and confounding variations to clarify the specifications. For example, safety testing may require harmful content examples and fairness testing may require demographic labels across groups.
-
Calculate required dataset sizes using statistical power analyses based upon the desired confidence level and interval for the criterion. Verify that the subgroup representation and sample sizes are adequate to test your release criteria with the required confidence you have set.
-
Consider whether one dataset can be used for multiple criteria. If so, verify that the statistical power offered by the dataset meets the needs of the most stringent release criterion.
-
Consider whether one criterion requires evaluation using multiple datasets. If your understanding of intrinsic and confounding variations is limited by known or unknown issues, your evaluation may benefit from using several independently sourced datasets.
Resources
Related documents:
