179 questions
No questions match those filters.
← All questions
EvaluationHard· Fragility· Log-Likelihood· Label-free Evaluation
You have no labeled evaluation for a new vertical. How...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansFragility is distinct from correctness. By creating variants of your prompt and measuring how much the model’s log-likelihood changes for the same output, you can quantify sensitivity without needing ground-truth labels. Decoding confidence can also serve as a useful proxy for monitoring fragility on live traffic where generating multiple variants is too expensive.