Minute Mysteries QA

A BIG-bench reading-comprehension task that asks a model, given a short crime story, to identify the perpetrator and explain the clues that support that deduction.

Also known as: MinuteMysteriesQA, minute mysteries

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
Subcategorynarrative mystery reading comprehension
Page statusunknown
Metricmultiple_choice_grade (MC subtask) and rougeLsum (free-text subtask); BIG-bench also plots a normalized aggregate
Directionhigher_is_better
Dataset size477
Dataset licenceBIG-bench repository Apache-2.0; Gutenberg stories under the Project Gutenberg License; 5minutemystery.com terms were not established (task authors said the site had no posted terms)
PublisherBIG-bench contributors (task authors listed below)

What it measures

Minute Mysteries QA gives the model a short crime or mystery story and asks who did it, and why that identification follows from the clues. Stories often hide a contradiction or a physically impossible claim in a suspect's account. The authors designed it to need causal, implicit and social reasoning, not factoid lookup. English text. Two subtasks share the stories: multiple choice (pick the perpetrator from a short list) and free text (name the culprit and the supporting reasoning).

Task format

JSON BIG-bench task with two subdirectories. Multiple choice: one correct perpetrator among three to five names, scored as multiple_choice_grade. Free text: a longer reference that includes the culprit and the explanation, scored with ROUGE and BLEU. Zero-shot in the task keywords.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub