List Functions

A BIG-bench task that infers one of 250 functions on lists of natural numbers from a few input/output pairs and applies it to a new list.

Also known as: BIG-bench list_functions

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 program induction over 250 list-of-integers functions
Page statusunknown
Metricexact_str_match
Directionhigher_is_better
Dataset size8000
Dataset licenceApache-2.0
PublisherGoogle (BIG-bench collaboration)

What it measures

list_functions is a program-induction probe, not a coding benchmark that asks for source. Each subtask names one function over lists of natural numbers (identity, drop ends, unique elements, cumulative sum, and 246 others). The model sees several input/output list pairs, then must emit the output list for a final input. Joshua S. Rule, Steven T. Piantadosi, and Joshua B. Tenenbaum built 250 hand-picked functions from Rule's 2020 MIT thesis so that human learners and symbolic program-induction models already have published per-function accuracies.

Task format

Free-text generation of a bracketed integer list. Preferred metric exact_str_match; also bleu and rouge. Parent task.json holds name and keywords; items live in c001–c250, 32 pairs each. Canary GUID embedded. Designed for zero-, one-, or many-shot use of those 32 pairs.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub