TruthfulQA: Measuring How Models Mimic Human Falsehoods

Authors: Lin et al. Year: 2022 ArXiv/Link: https://arxiv.org/abs/2109.07958

Summary

Benchmark measuring how often models give truthful answers when humans often give false ones.

Key Concepts

• Truthfulness
• Hallucination measurement
• Human biases
• Factuality
• Safety evaluation

Impact

Key benchmark for evaluating model truthfulness

Category

Safety & Evaluation

---