FActScore, proposed in 2023 at EMNLP by a research team from the University of Washington, University of Massachusetts Amherst, Allen Institute for AI, and Meta (Sewon Min et al.), is specifically designed to measure the factual precision of long-form text generation. It first decomposes generated content into independent atomic facts, then verifies each fact against reliable knowledge sources (such as Wikipedia) to determine if it is supported. The final score is the proportion of supported atomic facts out of the total atomic facts.

This metric focuses solely on precision and does not calculate recall. The score ranges from 0 to 100, typically reported as a percentage or a score between 0 and 1, where higher values are better, indicating a higher proportion of verifiable facts in the generated content. It is commonly used to evaluate the factual reliability of large language models when generating long-form texts such as biographies or encyclopedia entries.