LogiQA

LogiQA scores multiple-choice logical reasoning questions taken from China's National Civil Servants Examination, translated into English by professional translators.

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
Subcategorylogical reasoning reading comprehension, from Chinese civil-service exam questions
Page statusactive
Metricaccuracy (and length-normalized accuracy, acc_norm)
Directionhigher_is_better
Unit%
Dataset size8678
Dataset licenceNot stated by the paper or by the original GitHub repository, which carries no LICENSE file; the authors describe the dataset only as "freely available." The EleutherAI/logiqa Hugging Face mirror used by lm-evaluation-harness lists its own licence tag as "other" without further detail.
PublisherFudan University and Westlake University

What it measures

LogiQA tests deductive logical reasoning through reading comprehension: the model is given a short passage and a question, then must pick the correct one of four candidate answers. Questions come from "publically available questions of the National Civil Servants Examination of China," a real exam designed to test critical thinking rather than recall, and were not written for the benchmark. The paper organises items into five overlapping categories of deductive reasoning -- categorical, sufficient conditional, necessary conditional, disjunctive and conjunctive reasoning -- though a single question can draw on more than one category at once.

Task format

Four-option multiple-choice reading comprehension: a short context passage, a question about it, and four answer candidates, with the model returning a single choice. The dataset was released in both its original Chinese and a professionally produced English translation (five translators, three proofreaders), so the same questions can be evaluated in either language.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub