MMLU
A widely used 2020 benchmark of about 16,000 multiple-choice questions across 57 subjects.
Also: Massive Multitask Language Understanding
MMLU was published by Dan Hendrycks and colleagues in 2020. Its questions span school and university subjects, from elementary mathematics and US history to law, medicine and moral scenarios. GPT-3 scored about 44 per cent at release; frontier models passed 85 per cent by 2023 and approached 90 per cent by 2024, leaving little room to separate them. Studies have also found errors in some questions. Harder versions such as MMLU-Pro, and translated sets for other languages, including Indian ones, followed.