EusReading

352 Basque reading-comprehension items from EGA irakurmena papers (1998-2008), with long passages that the Latxa paper used as a 1-shot long-context probe.

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.
Categoryknowledge
SubcategoryBasque EGA irakurmena reading comprehension with long passages (1998-2008)
Page statusactive
Metricaccuracy
Directionhigher_is_better
Unit%
Dataset size352
Dataset licenceNot stated on the Hugging Face card (no licence tag). The Latxa paper says the evaluation datasets are "publicly available under open licenses" without naming one. The Latxa GitHub MIT licence is not confirmed as the dataset licence.
PublisherHiTZ Center - Ixa, University of the Basque Country (UPV/EHU)

What it measures

EusReading tests whether a model can read a long Basque passage and answer a multiple-choice question about it. The items are the irakurmena (reading) exercises from the same EGA C1 exam papers, 1998 to 2008, that supplied EusProficiency. The Latxa paper says each original test generally had about 10 questions, and that the passages are longer and harder than Belebele's Basque split, so they treat the set as a long-context reading probe rather than a short-span QA quiz.

Task format

Passage plus question plus two to four options in Basque. lm-evaluation-harness prefixes the passage with `Pasartea:`, then `Galdera:` and labelled options, ending with `Erantzuna:`. The paper evaluates it 1-shot because more exemplars would not fit most models' context windows, and scores by log-probability over the candidate letters.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub