ardge-hf/crag-qa

CRAG Task 3 (Parquet mirror)

Parquet re-packaging of Task 3 from Meta's CRAG: Comprehensive RAG Benchmark (KDD Cup 2024), split into two lightweight configs so the QA pairs can be loaded without pulling the full raw web content.

Source: data/crag_task_3_dev_v5.tar.bz2.part* in the upstream repo (2,706 QA pairs, each with 50 full-HTML web search results — the end-to-end RAG track: Task 1's 5-page retrieval plus Task 2's larger, noisier web result set combined).

Configs

• qa — one row per question: interaction_id, query_time, domain, question_type, static_or_dynamic, query, answer, alt_ans, popularity. No web content — a few hundred KB total.
• web — interaction_id + search_results (list of 50 structs: page_name, page_url, page_snippet, page_result (raw HTML), page_last_modified). Join to qa on interaction_id. ~8.8GB.

Both configs share the same validation (1,371 rows) / test (1,335 rows) split, mapped from the upstream split field (0 → validation, 1 → public test) so the two tables never mix across splits.

from datasets import load_dataset

qa = load_dataset("ardge-hf/crag-qa", "qa", split="validation")
web = load_dataset("ardge-hf/crag-qa", "web", split="validation")

Notes on the source data

• answer is stored as string for every row; a small number of upstream records carry a bare JSON number (e.g. 100) instead of a string, coerced here for a uniform schema.
• This mirror only covers Task 3. The upstream repo's Task 1/2 QA set uses the same 2,706 interactions with 5 pages per question instead of 50.

License

Original data licensed under CC BY-NC 4.0 by Meta Platforms, Inc. Non-commercial use only; retain attribution.

Citation

@article{yang2024crag,
      title={CRAG -- Comprehensive RAG Benchmark},
      author={Xiao Yang and Kai Sun and Hao Xin and Yushi Sun and Nikita Bhalla and Xiangsen Chen and Sajal Choudhary and Rongze Daniel Gui and Ziran Will Jiang and Ziyu Jiang and Lingkun Kong and Brian Moran and Jiaqi Wang and Yifan Ethan Xu and An Yan and Chenyu Yang and Eting Yuan and Hanwen Zha and Nan Tang and Lei Chen and Nicolas Scheffer and Yue Liu and Nirav Shah and Rakesh Wanga and Anuj Kumar and Wen-tau Yih and Xin Luna Dong},
      year={2024},
      journal={arXiv preprint arXiv:2406.04744},
      url={https://arxiv.org/abs/2406.04744}
}