BBC article-summary pairs across 45 language configs; OpenCompass concatenates validation splits and scores ROUGE.
unassessed
| Category | generation |
|---|---|
| Subcategory | multilingual abstractive summarization of BBC articles |
| Page status | active |
| Metric | ROUGE (OpenCompass RougeEvaluator) |
| Direction | higher_is_better |
| Dataset size | 1351253 |
| Dataset licence | CC-BY-NC-SA-4.0 (dataset contents remain BBC copyright; repo README restricts use to non-commercial research) |
| Publisher | Bangladesh University of Engineering and Technology (BUET) CSE NLP, with University of Rochester, Monash University, and Swinburne University |
XL-Sum asks a model to write a short abstractive summary of a BBC article. Articles and bullet-style summaries were extracted from BBC language sites with heuristics, not written by the authors. The v2 public release has about 1.35 million pairs. The ACL 2021 paper described an earlier cut of about 1 million pairs in 44 languages. OpenCompass loads every Hugging Face language config it lists and concatenates the validation splits, then scores ROUGE, so an OpenCompass "XLSum" number is a multilingual validation mix, not the English test split from the paper.
OpenCompass prompt: "Document：{text}\\nBased on the previous text, provide a brief single summary:". GenInferencer, ZeroRetriever (zero-shot). Predictions pass through Xsum_postprocess and RougeEvaluator. Hugging Face configs are named in full (english, chinese_simplified, …), not ISO codes.
No model card in ModelSpec reports this benchmark yet.