A 9,000-document BIG-bench task: label which sentences end paragraphs in nine European languages as a 0/1 sequence.
unassessed
| Category | generation |
|---|---|
| Subcategory | multilingual sentence-level paragraph-boundary labelling |
| Page status | unknown |
| Metric | exact_str_match |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 9000 |
| Dataset licence | Apache-2.0 |
| Publisher | Google (BIG-bench collaboration); task authors at Wrocław University of Science and Technology |
paragraph_segmentation feeds a sequence of consecutive sentences from a Wikipedia article and asks for a same-length sequence of 0/1 labels: 1 if the sentence ends a paragraph, 0 otherwise. Authors at Wrocław University of Science and Technology sampled 1,000 articles per language from Wikimedia dumps (tables, images, links, headings, and templates stripped; sentences split with Moses). Languages: German, Spanish, Italian, Russian, Polish, English, French, Dutch, Portuguese. The stated skill is detecting a semantic break, with speech formatting and short-context packing as intended uses.
Free-text label string. preferred_score exact_str_match. task_prefix tells the model to emit 0 and 1 in sentence order as one string. The worked example uses spaces ("0 0 1"). Canary GUID embedded. Dummy-model header: 0 multiple-choice and 9,000 free-text queries.
No model card in ModelSpec reports this benchmark yet.