GLUE: CoLA (Corpus of Linguistic Acceptability)

GLUE's single-sentence task: judge whether an English sentence is grammatically acceptable, scored by Matthews correlation because the acceptable/unacceptable classes are unbalanced.

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
Subcategorylinguistic acceptability judgement (single-sentence, English)
Page statussuperseded
MetricMatthews correlation coefficient (MCC)
Directionhigher_is_better
Unitcorrelation, -1 to 1
Dataset size1043
Dataset licenceDataset card licence: other; see the glue family page for the composite-licence explanation.
PublisherNew York University

What it measures

One English sentence, drawn from books and journal articles on linguistic theory, and one judgement: is it a grammatically acceptable sentence of English, or not. Unlike most of the rest of GLUE, this is a single-sentence task -- there is no second sentence to compare it against. About 70% of the training examples are labelled acceptable (6,023 of 8,551, per the Hugging Face datasets-server's own column statistics), unbalanced enough that raw accuracy would reward always guessing the majority class, which is why GLUE scores CoLA by Matthews correlation coefficient instead of accuracy.

Task format

Binary single-sentence classification (acceptable / unacceptable), commonly cast by harnesses as a two-way multiple-choice or yes/no generation task. Scored by Matthews correlation coefficient (MCC), a -1-to-1 correlation measure, not by accuracy.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub