Your question is Evaluating Beyond BLEU. Take a moment with it on the right.
Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).
How do you evaluate the performance of a model when standard metrics like BLEU or CHRF are insufficient?
Use this as a practical evaluation design question. Focus on how you would choose or build metrics, validate them against human judgment, and diagnose failure modes when n-gram overlap does not capture quality. Keep your answer concrete and structured.
