Evaluation
TruthfulQA
Quick Answer
A benchmark measuring whether models generate factually accurate answers or hallucinate.
TruthfulQA measures model honesty and factuality. It includes 817 questions where the natural language answer is often false. Evaluation uses GPT-3 + human review. TruthfulQA reveals hallucination tendencies. Fine-tuned models score better than base models. TruthfulQA shows instruction-tuning can increase truthfulness. It's important for measuring safety/alignment. TruthfulQA is increasingly standard for evaluation.
Last verified: 2026-04-08