{
  "preferred_language": "zh-en",
  "sources": [
    {
      "name": "DuReader",
      "language": "zh",
      "role": "primary Chinese QA/RAG source",
      "url": "https://github.com/baidu/DuReader",
      "paper": "https://arxiv.org/abs/1711.05073",
      "notes": "Real Chinese search questions with evidence documents and human answers. Check the upstream license before redistributing converted data."
    },
    {
      "name": "Natural Questions",
      "language": "en",
      "role": "open-domain QA source",
      "url": "https://github.com/google-research-datasets/natural-questions",
      "notes": "Google search questions with Wikipedia-based annotations. Useful for English QA and cross-lingual ablations."
    },
    {
      "name": "MS MARCO",
      "language": "en",
      "role": "retrieval and passage QA source",
      "url": "https://microsoft.github.io/msmarco/",
      "paper": "https://arxiv.org/abs/1611.09268",
      "notes": "Bing search queries, passages, and answers. Useful for hard negative construction and retrieval-style evaluation."
    },
    {
      "name": "KILT",
      "language": "en",
      "role": "provenance-oriented evaluation source",
      "url": "https://github.com/facebookresearch/KILT",
      "paper": "https://arxiv.org/abs/2009.02252",
      "notes": "Knowledge-intensive benchmark with provenance. Useful for citation/provenance metrics."
    }
  ],
  "derived_format": {
    "name": "SearchShortQA v1",
    "required_fields": ["id", "question", "answer", "answers", "gold_doc_ids", "contexts"],
    "recommended_splits": {
      "sft_train": "30000-100000",
      "grpo_train": "3000-10000",
      "dev": "1000-3000",
      "test": "1000-3000"
    }
  }
}