> ML_DATASET // TRUTHFULQA-DATASET_v1.0
TruthfulQA: Measuring Model Imitation of Falsehoods
Oxford / Evans Lab (Lin et al. 2022) · Epistemic Truthfulness & Alignment · 817 adversarial questions across 38 categories (health, law, conspiracies, superstitions)
Epistemic Truthfulness & AlignmentApache-2.0817 adversarial questions across 38 categories (health, law, conspiracies, superstitions)open
Dataset Profile & Characteristics
Label Type:Multiple-choice and generation reference true/false answers
Languages:en
License Tier:permissive-open-source
Modalities:text
Intended Use
- Evaluating model truthfulness and hallucination susceptibility in high-stakes domains
Prohibited / Discouraged Use
- Evaluating general factual trivia recall capacity
Bias, Leakage & Privacy Risk Analysis
Privacy / Sensitive Data Risks:
Questions designed to mimic common human cognitive fallacies; zero personal data.
Known Bias:
Targeted adversarial design intentionally provoking superstitious falsehoods.
Known Benchmark Leakage:
Benchmark de-contamination needed when fine-tuning alignment models.