SuperGLUE (Super General Language Understanding Evaluation benchmark)

An eight-task English language-understanding suite that replaced GLUE once models exceeded its human baseline, itself since superseded by harder generative and agentic benchmarks.

Also known as: Super General Language Understanding Evaluation benchmark, SGLUE

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.
Categorycomposite
Subcategorymulti-task English sentence and sentence-pair understanding suite, successor to GLUE
Page statussuperseded
MetricSuperGLUE Score: unweighted average of eight per-task scores (accuracy for five tasks; accuracy and F1 averaged for CB; F1a and exact match averaged for MultiRC; F1 and exact match averaged for ReCoRD)
Directionhigher_is_better
Unitpoints
Dataset size25730
Dataset licenceThe Hugging Face dataset card lists licence "other": SuperGLUE reuses or extends multiple pre-existing datasets (CommitmentBank, COPA, MultiRC, ReCoRD, RTE, WiC, the Winograd Schema Challenge, Winogender) and each retains its own original terms rather than one unified SuperGLUE licence.
PublisherNew York University

What it measures

SuperGLUE bundles eight separately scored English understanding tasks behind one composite score, chosen specifically to be harder than GLUE's tasks and to cover a wider range of task formats: question answering with yes/no answers over a passage (BoolQ), three-way textual entailment from naturally occurring discourse (CB, CommitmentBank), causal reasoning between two alternatives (COPA), multi-sentence reading comprehension with a variable number of correct answers (MultiRC), span-based reading comprehension requiring cloze-style entity prediction (ReCoRD), two-way textual entailment (RTE), word-sense disambiguation in context (WiC), and coreference resolution via the Winograd schema (WSC). Two further diagnostic sets, AX-b (broad-coverage entailment) and AX-g (Winogender gender-bias probe), are reported alongside the suite but excluded from its main score. All tasks are English text classification or span extraction, not open-ended generation.

Task format

Eight separately scored tasks, each single-sentence, sentence-pair, or short-passage classification or span extraction; each keeps its own metric (see How it is scored) and the eight per-task scores are averaged, unweighted, into one SuperGLUE Score out of 100. Two diagnostic sets (AX-b, AX-g) are reported separately and not folded into the score.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub