hotchpotch/wikipedia-multilingual-ir-pairs
wikipedia-multilingual-ir-pairs This dataset is designed for multilingual information retrieval training. Compared with raw Wikipedia paragraph dumps, it provides cleaner and more practical supervision by pairing title/section-driven queries with relevant paragraph-level documents and applying rule-based filtering to remove low-value sections and noisy fragments. It is intended for IR model training, including contrastive learning and retrieval/reranking objectives.… See the full description on the dataset page: https://huggingface.co/datasets/hotchpotch/wikipedia-multilingual-ir-pairs.
wikipedia-multilingual-ir-pairs
This dataset is designed for multilingual information retrieval training. Compared with raw Wikipedia paragraph dumps, it provides cleaner and more practical supervision by pairing title/section-driven queries with relevant paragraph-level documents and applying rule-based filtering to remove low-value sections and noisy fragments. It is intended for IR model training, including contrastive learning and retrieval/reranking objectives.
Dataset at a glance
- Task: IR training (retrieval and reranking)
- Fields:
query(Wikipedia title, or title + section header),document(paragraph text from the corresponding article section) - Example query construction: if title is
Kyotoand section header isHistory, the query becomesKyoto Historyand is paired with the corresponding section text. - Total rows: 95,757,423
- Language subsets: 11
- Source dataset: singletongue/wikipedia-paragraphs
Language Subsets
Dataset Creation Process (Rough)
The pipeline is implemented in `scripts/build_subheader_filtered_hf_ds.py`.
- Load each latest language subset from
singletongue/wikipedia-paragraphs. - Filter out non-article pages and low-value sections (for example, "See also", references, external links; language-specific rules).
- Apply additional quality filters (short/noisy list fragment removal, disambiguation lead gates, query-length constraints, and dropping pairs where
len(query) > len(document)). - Split and merge paragraphs into practical chunks (character-length targets: 1000 for non-CJK, 600 for Japanese/Korean, 500 for Chinese; newline-aware splitting with an optional split cap; balanced merge for consecutive paragraphs in the same section; and a final hard cap on document length).
- Save each language subset with
query/documentcolumns.
Acknowledgements
Thank you for publishing the very useful dataset singletongue/wikipedia-paragraphs. This dataset made it possible to build this IR-focused dataset quickly.
License
This dataset is derived from Wikipedia text and follows Wikipedia licensing:
- Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)
- GNU Free Documentation License (GFDL), where applicable
Downstream users should preserve attribution and share-alike obligations for the Wikipedia-derived text.
