RACE

RACE is multiple-choice reading comprehension from English exams written for Chinese middle- and high-school students, split into easier RACE-M and harder RACE-H passages, four options per question.

Also known as: ReAding Comprehension dataset from Examinations

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.
Categoryreasoning
Subcategorymultiple-choice reading comprehension: English-exam questions written for Chinese middle- and high-school students
Page statusactive
MetricAccuracy (percentage of questions answered correctly)
Directionhigher_is_better
Unit%
Dataset size97687
Dataset licenceNon-commercial research use only (custom licence; Hugging Face tags it 'other' rather than a standard SPDX identifier)
PublisherCarnegie Mellon University

What it measures

RACE gives a model a passage taken from an English-language exam administered to Chinese students aged roughly 12 to 18, together with a question about it and four answer options, exactly as the question appeared on the real exam. Questions come in two forms: ordinary interrogative questions, and cloze-style questions with an underscore standing in for a missing phrase that the correct option must complete. The passages were written by English teachers specifically to test the students' reading comprehension and reasoning, not sampled from generic web text, so RACE leans more on inference, vocabulary and paraphrase than on locating a matching sentence the way span-extraction sets like SQuAD (`squad`) do. The dataset is split by school level into RACE-M (middle school, easier) and RACE-H (high school, harder); a combined "all" configuration pools both.

Task format

Four-way multiple choice: given a passage and a question (or cloze sentence), select the correct one of four lettered options (A-D). Harness implementations vary in whether they present this as a loglikelihood ranking over the four options or a generation task asking the model to output a letter.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub