{
  "schema": "silktrain-pilot-sample/v1",
  "synthetic": true,
  "locale": "zh-hans",
  "workflow": "preference-data",
  "service": "training-evaluation",
  "title": "偏好排序与 RLHF 数据",
  "source_scenario": "两条回答概述同一政策，其中一条添加了没有依据的例外。",
  "expected_delivery": "用于 DPO 或奖励模型训练的回答对、优选与弃选标记、理由及平局记录。",
  "acceptance_check": "用同一量表双人评审抽样记录，导出前裁决分歧。",
  "scoping_checklist": "准备一小组回答、参考来源、评分准则，以及需要第二人复核的示例。",
  "examples": [
    {
      "id": "P-023",
      "source": "Orders can be cancelled before dispatch.",
      "a": "All orders can be cancelled.",
      "b": "Cancellation is available before dispatch.",
      "chosen": "B",
      "rejected": "A",
      "criterion": "source_fidelity"
    },
    {
      "id": "P-024",
      "source": "Support replies within two working days.",
      "a": "Expect a reply within two working days.",
      "b": "Expect a reply within 48 hours.",
      "chosen": "A",
      "rejected": "B",
      "criterion": "business_days"
    },
    {
      "id": "P-025",
      "source": "Use reference K9 for this example.",
      "a": "The reference is K9.",
      "b": "Your reference is K9.",
      "chosen": null,
      "rejected": null,
      "tie": true,
      "criterion": "equivalent"
    }
  ]
}