SCALABLE WORKFLOW FOR EVALUATING TRUSTWORTHINESS OF LARGE LANGUAGE MODELS

Summary

Large Language Models (LLMs) present significant trustworthiness challenges, including generating inaccurate or biased outputs and potentially misaligning with human values. Oak Ridge National Laboratory has developed a scalable workflow to comprehensively evaluate open-source LLMs across multiple dimensions of trustworthiness, such as truthfulness, safety, and fairness. This workflow has been applied to evaluate Meta Llama3.1 LLMs and aims to lay the groundwork for future development of more trustworthy AI models.

Key Facts

Source Document

https://example-government.gov/policy-document-link