Linguistics Puzzles

A 2,000-item BIG-bench Lite task that induces a constructed language from paired sentences and translates one more sentence either way.

Also known as: BIG-bench linguistics_puzzles, Linguistic Puzzles

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 Lite Rosetta-stone puzzles over constructed languages (2,000 items)
Page statusunknown
Metricexact_str_match
Directionhigher_is_better
Dataset size2000
Dataset licenceApache-2.0
PublisherGoogle (BIG-bench collaboration)

What it measures

linguistics_puzzles is a Rosetta-stone linguistics-olympiad item written over a fresh constructed language, not a natural low-resource language. Each example shows several English/conlang sentence pairs, then asks for one more translation either into English or into the conlang. Nathan A. Chi generated a different conlang per item (case, number, headedness, and similar features) so web lookup of a real language should not help. It is not [LingOly](lingoly.md) (real olympiad sheets) and not [conlang_translation](conlang_translation.md) (ciphered real languages on BIG-bench Lite).

Task format

Free-text generation. Preferred metric exact_str_match; also bleu and rouge. One parent task.json holds all 2,000 examples. Canary GUID embedded. BIG-bench Lite member.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub