chrischarts8/pcare-dim4-phase2-synthetic

PCare DIM4 Phase 2 - Clinical Outcome Comparator (Synthetic Training Data)

Overview

Synthetic dataset for training a clinical outcome comparator model. Each example contains an expected clinical outcome (from document analysis) and an actual clinical outcome (from a compiled FHIR artifact), and the model must judge whether they are semantically equivalent.

Task

Input: Two clinical outcome strings Output: JSON with verdict (PASS/FAIL), reasoning, and mismatches

Dataset Statistics

Split	Total	PASS	FAIL	Balance
Train	1,196	592	604	49.5% / 50.5%
Eval	299	157	142	52.5% / 47.5%

Scenario Types

PASS Scenarios (7 types)

• P1: Exact semantic match, different wording (120)
• P2: Actual more specific but clinically consistent (100)
• P3: Expected more detailed but clinically consistent (100)
• P4: Negative/exclusion equivalence (100)
• P5: Category/grade mapping (100)
• P6: Extra domains in actual (80)
• P7: Format difference only (150)

FAIL Scenarios (8 types)

• F1: Wrong clinical action (120)
• F2: Wrong patient population (100)
• F3: Missing outcome domain (100)
• F4: Wrong severity/grade (80)
• F5: Opposite action (100)
• F6: NO ACTION MATCHED system error (80)
• F7: Wrong diagnostic classification (80)
• F8: Partial match with critical omission (90)

Clinical Domains

1. Cancer screening (colorectal, breast, cervical, lung, prostate)
2. Diagnostic classification (transfusion reactions, sepsis, stroke)
3. Treatment guidelines (antibiotics, anticoagulation, pain management)
4. Risk stratification (cardiovascular, fall risk, nutrition)
5. Immunization schedules (childhood, adult, travel)
6. Preventive care (wellness visit components)
7. Referral criteria (specialist referrals)
8. Monitoring protocols (lab monitoring, imaging follow-up)

Format

Messages format compatible with TRL SFTTrainer with assistant_only_loss=True:

{
  "messages": [
    {"role": "system", "content": "You are a clinical outcome comparator..."},
    {"role": "user", "content": "## Expected Outcome...\n\n## Actual Outcome..."},
    {"role": "assistant", "content": "{\"verdict\": \"PASS\", \"reasoning\": \"...\", \"mismatches\": []}"}
  ]
}