Also known as: Reading Comprehension With Commonsense Reasoning Dataset (ReCoRD)
Reading Comprehension with Commonsense Reasoning Dataset, или ReCoRD, — benchmark для оценки способности языковой модели использовать commonsense reasoning при ответах на вопросы с замаскированными сущностями из текстовых фрагментов. Он входит в набор benchmark-задач SuperGLUE и проверяет, может ли модель вывести правильную сущность из контекстных доказательств.
A dataset to evaluate an LLM's ability to perform commonsense reasoning. Each example in the dataset contains three components: - A paragraph or two from a news article - A query in which one of the entities explicitly or implicitly identified in the passage is masked. - The answer (the name of the entity that belongs in the mask) See ReCoRD for an extensive list of examples. ReCoRD is a component of the SuperGLUE ensemble.