SuperGLUE ReCoRD (Reading Comprehension with Commonsense Reasoning Dataset)

SuperGLUE's cloze reading-comprehension task: recover a masked entity in a CNN/Daily Mail query, scored by token-level F1 and exact match.

Also known as: ReCoRD, SuperGLUE_ReCoRD, record

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
Subcategorycloze reading comprehension over news, with a masked named entity
Page statussaturated
Metricmax token-level F1 and exact match (EM), averaged for the SuperGLUE task score
Directionhigher_is_better
Unit%
Dataset size10000
Dataset licenceHugging Face card licence is "other". The ReCoRD site says CNN/Daily Mail passages follow the Apache licence of the DeepMind rc-data dump, and Internet Archive crawls follow that archive's terms of use.
PublisherNew York University (SuperGLUE); original dataset from Johns Hopkins University and Microsoft

What it measures

ReCoRD asks a model to read a news passage and a cloze query in which one entity is replaced by a blank, then name the missing entity. Candidate answers are entities that already appear in the passage; several surface forms of the same entity all count as correct. The items come from CNN and Daily Mail articles. The original authors designed the queries so that many of them need commonsense about the article, not only span matching. SuperGLUE reports the task as English text-only reading comprehension and folds F1 and exact match into the suite average.

Task format

Cloze-style entity recovery given a passage, a query with a blank, and a list of in-passage entities. SuperGLUE and lm-evaluation-harness treat it as multiple-choice over those entities. OpenCompass instead generates an entity name and scores exact match against the gold aliases.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub