Data Wrangling (BIG-bench)

A 246-file BIG-bench suite of string wrangling problems: infer a format change from few-shot pairs and emit the transformed string.

Also known as: mult_data_wrangling, BIG-bench Data Wrangling

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 few-shot string-format transformation (dates, emails, names, phones, times, units, free text)
Page statusunknown
Metricexact_str_match
Directionhigher_is_better
Unit%
Dataset size7872
Dataset licenceApache-2.0
PublisherUniversitat Politècnica de València (BIG-bench collaboration)

What it measures

mult_data_wrangling asks a model to finish a string transformation after seeing one or more input-output pairs from the same problem. Domains are dates, emails, free text, names, phones, times, and units (for example turning a compact date into hyphenated day-month-year, or taking the domain of an email). Authors at Universitat Politècnica de València built 123 problems and shipped each in two prompt styles (input-output and a short English statement). The skill is few-shot format induction on English-script strings, not SQL or table joins. It is not [entity_data_imputation](entity_data_imputation.md).

Task format

Free-text completion scored with exact_str_match. example_input_prefix "Input: ", example_output_prefix " Output: ", output_regex ("'.*?'") on the input-output style files. Keyword many-shot. Canary GUID embedded. Parent task.json has no examples; they live in 246 subtask files.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub