TruthfulQA

A benchmark of 817 questions designed to test whether language models generate truthful answers, specifically targeting common misconceptions and falsehoods that models tend to reproduce.