unsupported-claim rate
The percentage of claims in a response that aren't grounded.
Plain English Explanation
The percentage of claims in a response that aren't grounded. For example, if an LLM's response makes 10 claims but only 1 is grounded, the UCR is 90%. A high UCR implies that an LLM is hallucinating too frequently. See also citation precision and citation recall.
How is it used?
Practitioners refer to unsupported-claim rate when building, training, or evaluating machine learning systems. It appears in research papers, product documentation, and technical discussions about AI capabilities and limitations.