PortugueseBench

IberoBench's European Portuguese suite: ASSIN entailment and paraphrase, Belebele reading, and FLORES translation, plus two extra ASSIN2 tasks in the current harness group.

Also known as: portuguese_bench, Portuguese Bench

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.
Categorycomposite
SubcategoryEuropean Portuguese IberoBench slice: NLI, paraphrase, reading comprehension and translation
Page statusactive
Metrictask-dependent: accuracy (ASSIN, Belebele); BLEU/ROUGE-family scores for FLORES; macro-F1 (assin2_rte); Pearson r (assin2_sts)
Directionhigher_is_better
Dataset licenceVaries by component. facebook/belebele is CC-BY-SA-4.0. facebook/flores is CC-BY-SA-4.0 (gated). nilc-nlp/assin and nilc-nlp/assin2 list licence as unknown on their Hugging Face cards. No suite-wide licence is established.
PublisherIberoBench consortium: Barcelona Supercomputing Center (BSC-CNS), Universitat Pompeu Fabra, CiTIUS-USC, and HiTZ Center - IXA (UPV/EHU)

What it measures

PortugueseBench is the European Portuguese slice of IberoBench, the same five-language project as this repository's CatalanBench, BasqueBench and GalicianBench pages. IberoBench Table 1 lists four Portuguese tasks: assin_entailment (NLI), assin_paraphrase, belebele_por_Latn (reading comprehension) and flores_pt (translation). The paper states that several task types are not yet available for European Portuguese. Translations into Portuguese were aimed at the European variety. The current lm-evaluation-harness group also runs assin2_rte and assin2_sts, which Table 1 does not list.

Task format

Mixed: two-way multiple choice for ASSIN entailment and paraphrase; multiple-choice reading for Belebele Portuguese; sentence-level machine translation for sixteen FLORES directions to or from Portuguese; multiple-choice RTE and free-form 1-5 similarity for ASSIN2 in the harness-only extras. IberoBench evaluates 0-shot and 5-shot.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub