How to Choose a Threshold for an Evaluation Metric for Large Language Models