{
  "schema": "silktrain-pilot-sample/v1",
  "synthetic": true,
  "locale": "en",
  "workflow": "preference-data",
  "service": "training-evaluation",
  "title": "Preference ranking and RLHF data",
  "source_scenario": "Two answers summarize the same policy. One adds an unsupported exception.",
  "expected_delivery": "Ranked response pairs with the chosen and rejected answer, a rationale and ties for DPO or reward-model training.",
  "acceptance_check": "Double-score a sample against the rubric and resolve reviewer disagreement before export.",
  "scoping_checklist": "Prepare a small answer set, the source material, a scoring rubric and examples that should go to a second reviewer.",
  "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"
    }
  ]
}