TruthfulQA

817 questions written to bait a model into repeating common human misconceptions, testing truthfulness rather than raw knowledge.

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorysafety
Subcategorytruthfulness / imitative falsehood
Page statusunknown
MetricMC1 / MC2 accuracy, or GPT-judge truthful-and-informative rate
Directionhigher_is_better
Unit%
Dataset size817
Dataset licenceApache-2.0
Publishernot established (academic collaboration)

What it measures

TruthfulQA gives a model 817 questions across 38 categories - including health, law, finance, politics, and topics like conspiracies and superstitions - that were specifically written because some humans would answer them falsely due to a popular misconception. It measures whether a model repeats that popular falsehood or gives the truthful answer, which the original paper distinguishes from ordinary factual-knowledge testing: a model can "know" the correct fact internally and still be more likely to output the popular wrong answer, because that is the pattern most rewarded in its training text.

Task format

Multiple-choice (pick the true statement from several) and free-form generation, over 817 fixed questions across 38 categories.

Models reporting this benchmark

These figures come from the model cards, which carry one collection date per card and no per-score attribution. They are shown as reported, not as verified evidence.
ModelProviderScoreCard as of
Mistral 7B Instruct v0.2Mistral AI68.32024-07
Mixtral 8x22B Instruct v0.1Mistral AI68.12026-04
Mixtral 8x7B Instruct v0.1Mistral AI65.02026-04
Yi 1.5 34B Chat01.AI62.22026-04
Yi 1.5 34B Chat 16K01.AI62.22024-07
Meta Llama 3 70BMeta61.82024-07
Meta Llama 3 70B InstructMeta61.82026-04
Meta Llama 3 70B InstructNous Research61.82026-04
Nous Hermes 2 Yi 34BNous Research60.42024-07
Phi 3 mini 4K instructMicrosoft59.92024-07
Hermes 2 Pro Llama 3 8BNous Research56.62024-07
Nous Hermes 2 SOLAR 10.7BNous Research55.82024-07
Hermes 2 Theta Llama 3 8BNous Research55.72024-07
Yi 34B Chat01.AI55.42024-07
Nous Hermes 2 Mixtral 8x7B DPONous Research54.82024-07
Phi 3 mini 128K instructMicrosoft54.12024-07
Yi 1.5 34B01.AI53.82026-04
Yi 1.5 9B Chat01.AI52.72024-07
Yi 1.5 6B Chat01.AI52.62024-07
Yi 1.5 34B 32K01.AI52.12024-07
Meta Llama 3 8B InstructMeta51.62024-07
Meta Llama 3 8B InstructNous Research51.62024-07
Yi 1.5 9B Chat 16K01.AI51.02024-07
chatglm2 6BZhipu AI48.12024-07
deepseek llm 7B baseDeepSeek47.92024-07
deepseek llm 7B chatDeepSeek47.92024-07
Mixtral 8x7B v0.1Mistral AI46.82026-04
Yi 1.5 9B01.AI46.72024-07
Qwen2 1.5B InstructAlibaba / Qwen Team45.92024-07
gemma 2B itGoogle DeepMind45.82024-07
deepseek coder 6.7B instructDeepSeek45.62024-07
gemma 7B itGoogle DeepMind44.92024-07
phi 2Microsoft44.22024-07
deepseek coder 1.3B baseDeepSeek44.02024-07
deepseek coder 1.3B instructDeepSeek44.02024-07
Yi 1.5 6B01.AI44.02024-07
Meta Llama 3 8BMeta42.92024-07
Meta Llama 3 8BNous Research42.92024-07
Yi 34B 200K01.AI42.62024-07
Yi 1.5 9B 32K01.AI42.52024-07
Yi 9B01.AI42.42024-07
Yi 6B01.AI42.02024-07
Yi 6B Chat01.AI42.02024-07
Mistral 7B v0.3Mistral AI41.82024-07
mistral 7B v0.3 bnb 4bitUnsloth41.82024-07
falcon 40BTII41.72024-07
deepseek coder 6.7B baseDeepSeek40.32024-07
Qwen2 0.5B InstructAlibaba / Qwen Team39.82024-07
gemma 2BGoogle DeepMind33.12024-07
OLMo 1B hfAllen AI32.92024-07

Data

This page as JSON · Edit on GitHub