Similarities Test for Abstraction

A 76-item BIG-bench similarities test that asks how two objects are alike, scoring the abstract option over concrete distractors.

Also known as: Similarities Test for Abstraction

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
SubcategoryBIG-bench MoCA/WAIS-style object similarities, abstract vs concrete (76 items)
Page statusunknown
Metricmultiple_choice_grade
Directionhigher_is_better
Unit%
Dataset size76
Dataset licenceApache-2.0
PublisherGoogle (BIG-bench collaboration)

What it measures

similarities_abstraction names two objects and asks how they are alike, after a fruit example in task_prefix, following MoCA/WAIS similarities practice. M. Yee wrote one abstract gold and several concrete distractors per item (truthful but too specific). Direct count: 76 examples (66 with four choices, 7 with five, 3 with six). Preferred metric multiple_choice_grade; BLEU and ROUGE are also listed for optional free response. Dummy-model header: 76 multiple-choice and 76 free-text queries. Canary GUID embedded. Not in BIG-Bench Hard.

Task format

Multiple choice with append_choices_to_input false, plus optional free-response targets for BLEU and ROUGE. JSON task. Keywords analogical reasoning, context-free question answering, free response, human-like behavior, json, multiple choice.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub