prompts/syn_dataset_subclaims_support_check.txt

You are an expert in biomedical NLP and clinical evidence reasoning.
Your task is to generate synthetic medical data for training a model that determines whether a given long medical text supports a subclaim.

For each dataset item:

1. Create **one medical text** (6–10 sentences).
2. Create **12 atomic subclaims** about the text.
3. Assign each subclaim a label:

   * `"supported"` → The text directly supports the subclaim.
   * `"refuted"` → The text contradicts the subclaim.
   * `"not_supported"` → The text is related but has no evidence.

Requirements:

* All content must be **synthetic**, **plausible**, and **medically coherent**.
* Subclaims must be **short** and **atomic** (only one fact).
* Keep wording efficient to reduce tokens.
* Ensure diversity across diseases, patient populations, treatments, and outcomes.
* Make labels unambiguous.

Return output **strictly** in JSON.


Generate **2 dataset items**.
For each item:

* Create **one 6–10 sentence medical text** about a clinical condition, treatment, diagnostic method, or patient group.
* Then create **12 subclaims**, labeled:

  * 4 `"supported"`
  * 4 `"refuted"`
  * 4 `"not_supported"`

Use the JSON structure exactly:

```json
{
  "items": [
    {
      "text": "TEXT_1",
      "subclaims": [
        {"subclaim": "…", "label": "supported"},
        {"subclaim": "…", "label": "supported"},
        {"subclaim": "…", "label": "supported"},
        {"subclaim": "…", "label": "supported"},
        {"subclaim": "…", "label": "refuted"},
        {"subclaim": "…", "label": "refuted"},
        {"subclaim": "…", "label": "refuted"},
        {"subclaim": "…", "label": "refuted"},
        {"subclaim": "…", "label": "not_supported"},
        {"subclaim": "…", "label": "not_supported"},
        {"subclaim": "…", "label": "not_supported"},
        {"subclaim": "…", "label": "not_supported"}
      ]
    }
  ]
}
```
Generate **2 such items**.