OpenBookQA

5,957 four-way elementary-science questions built around a small fact book; human accuracy is near 92%, and no current model card or live leaderboard still reports it.

unverified

This page is not in the default catalogue. Evidence required by the catalogue contract is missing or was not approved by a reviewer. That is a statement about the evidence we hold, not a claim that the benchmark is stale or illegitimate.

Recorded reasons:

Categoryreasoning
Subcategoryopen-book elementary-science multiple-choice QA
Page statussaturated
Metricaccuracy (acc and length-normalised acc_norm)
Directionhigher_is_better
Unit%
Dataset size5957
Dataset licenceMarked "unknown" on the Hugging Face dataset card. The reference code repository (allenai/OpenBookQA on GitHub) is Apache-2.0 licensed, but that licence covers the baseline modelling code, not a separately stated licence for the question-and-fact text itself.
PublisherAllen Institute for Artificial Intelligence (AI2)

What it measures

OpenBookQA asks a model to answer an elementary-school-level science question by combining one "open book" fact with broad common knowledge, the way an open-book exam works: knowing where to look up the fact is not enough, because the missing piece is background knowledge no book states outright. The open book itself is a set of roughly 1,326-1,329 core science facts (sources read for this page give slightly different counts; see Dataset and licence) covering topics such as conductivity, states of matter and food chains. Around 6,000 four-way multiple-choice questions were then written against those facts so that each one also requires an additional, unstated piece of common knowledge to connect the fact to the question -- for example, the fact "metal is a thermal conductor" only answers a question about which object conducts the most heat if the model also knows a steel spoon is made of metal. That two-hop structure sets it apart from earlier reading-comprehension datasets, which typically expect an answer extractable from one supplied passage.

Task format

Four-way multiple-choice question answering, typically zero- or few-shot, over a fixed set of roughly 6,000 English-language elementary-science questions; no passage or fact is supplied at inference time, so the model must supply both the fact and the connecting common-knowledge step itself.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub