This service analyzes the model against a diverse set of Out-of-Distribution (OOD) examples. These examples include prompts whose content the model did not see during its training. Understanding the model’s behavior when faced with OOD examples is important because the model may behave unpredictably when seeing unexpected content, leading to potential safety concerns. The OOD Robustness score aggregates the scores across the different tests under this service, where each test employs a distinct dataset of OOD examples that are obtained using various text transformation methods. This score is an indicator of the model’s ability to understand and correctly act on different types of prompts that it hasn’t previously seen.