ORIGINAL TECHNICAL PRACTICE
Build a generative-answer evaluation set
What would you include in an evaluation set before releasing an AI assistant that answers internal product questions?
Written for independent practice. The model answer presents one approach; clarify constraints and explain alternatives.
A REASONED APPROACH
Model answer.
Evaluate the behavior the assistant must deliver, including unsupported requests and unsafe instructions. Use a versioned set with evidence-backed expectations and a clear separation between development examples and release checks.
Work through it
- Sample realistic tasks across document types, question difficulty, user permissions and languages that are actually supported. Include ambiguous and unanswerable questions.
- Record acceptable answer elements, required evidence, unacceptable claims and when clarification or abstention is expected. Prefer testable criteria over a single stylistic reference answer.
- Measure task success, unsupported claims, source support, latency and cost. Audit a sample of automated grades and compare revisions with the same frozen inputs.
Tradeoffs to discuss
- Human review is costly but useful for nuanced failure modes. Model-based grading can expand coverage while introducing its own bias and nondeterminism.
Common pitfalls
- Training or prompt-tuning on the release set makes it less informative. A high aggregate score can conceal a severe failure in a small but important task category.