Upload folder using huggingface_hub

README.md

@@ -97,7 +97,7 @@ file_report()        → KARS PASSED, reward=+15
 ```bash
 # Start the server
 conda activate openenv
-uvicorn server.app:app --host 0.0.0.0 --port 8000
+uvicorn server.app:app --host 0.0.0.0 --port 7860
 ```
 
 ```python
@@ -120,11 +120,27 @@ trajectories = run_episodes(
 
 ## Patients
 
-| ID | Condition | T1 GFR | T5 GFR | HbA1c T1→T5 | Notes |
-|----|-----------|--------|--------|-------------|-------|
-| P001 | CKD Stage 4 | 18.5 | 12.1 | 7.2→8.9 | Complete record |
-| P002 | Diabetic nephropathy | 11.0 | 8.3 | 9.1→10.2 | Antihypertensives, insulin |
-| P003 | CKD Stage 3 | 22.3 | 19.8 | null | HbA1c never recorded, inactive waitlist |
+50 procedurally generated patients (P001–P050) across CKD stages 3–5:
+
+| ID | Condition | Notes |
+|----|-----------|-------|
+| P001 | CKD Stage 4 | Complete record, anchor patient |
+| P002 | Diabetic nephropathy | Antihypertensives, insulin, anchor patient |
+| P003 | CKD Stage 3 | HbA1c not recorded (non-diabetic), anchor patient |
+| P004–P050 | CKD Stage 3/4/5 | Procedurally generated (seed=42) |
+
+**Patient distribution:**
+- CKD Stage 3: ~25% · Stage 4: ~50% · Stage 5: ~25%
+- 60% diabetic — HbA1c present; non-diabetics have 85% chance of missing HbA1c
+- ~10% of patients have an injected anomalous lab reading (>25% delta) for benchmark coverage
+
+**All patients include distractor fields** (queryable but not KARS-required):
+`cholesterol`, `bmi`, `albumin`, `hemoglobin`
+
+To regenerate the patient database:
+```bash
+python data/generate_patients.py
+```
 
 ## KARS Required Fields
 
@@ -143,10 +159,12 @@ prana_env/
 ├── models.py                      # PranaAction, PranaObservation
 ├── test_agent.py                  # LLM agent RL loop (GPT-4o)
 ├── test_client.py                 # Smoke test client
+├── prana_grpo_qwen3_8b_fp8.ipynb  # GRPO fine-tuning notebook (Qwen3-8B FP8, H100)
 ├── data/
-│   └── patient_db.json            # Patient records with T1 snapshots and T5 values
+│   ├── patient_db.json            # 50 patients with T1 snapshots, T5 values, distractor fields
+│   └── generate_patients.py      # Procedural patient generator (seed=42, CKD stage distributions)
 └── server/
-    ├── app.py                     # FastAPI + WebSocket server
+    ├── app.py                     # FastAPI + WebSocket server (port 7860)
     ├── prana_env_environment.py   # RL environment: actions, KARS validator, rewards
     └── Dockerfile
 ```
@@ -157,7 +175,7 @@ prana_env/
 from prana_env.client import PranaEnv
 from prana_env.models import PranaAction
 
-with PranaEnv(base_url="http://localhost:8000") as env:
+with PranaEnv(base_url="http://localhost:7860") as env:
     result = env.reset(patient_id="P001")
     print(result.observation.query_result)
 

data/generate_patients.py

@@ -0,0 +1,238 @@
+#!/usr/bin/env python3
+"""
+Generate procedural patient database for PRANA-Env and tau2 benchmark.
+
+Produces 50 patients (P001-P050) across CKD stages 3-5 with:
+  - Stage-appropriate GFR / creatinine distributions
+  - Diabetic status driving HbA1c presence
+  - Systematic missing fields (non-diabetic → null HbA1c, etc.)
+  - Distractor fields (cholesterol, bmi, albumin, hemoglobin)
+  - T1 snapshot with slightly better values (disease progression)
+  - Anomaly injection for ~10% of patients (for benchmark coverage)
+
+Outputs:
+  - prana_env/data/patient_db.json        (prana_env flat format)
+  - tau2-bench/data/tau2/domains/prana/db.json  (tau2 LabResult format, preserves other DBs)
+"""
+
+import json
+import random
+from pathlib import Path
+from datetime import date, timedelta
+
+SEED = 42
+N_PATIENTS = 50
+EPISODE_DATE = date(2026, 3, 7)
+T1_NOMINAL_DATE = date(2025, 11, 7)
+T5_MEAS_DATE = date(2026, 3, 1)      # recent measurement date used in tau2 history
+
+rng = random.Random(SEED)
+
+# ── Clinical distributions ─────────────────────────────────────────────────────
+
+CKD_STAGES = {
+    3: {"gfr": (30, 59), "creatinine": (1.2, 2.5), "weight": 0.25},
+    4: {"gfr": (15, 29), "creatinine": (2.5, 5.0), "weight": 0.50},
+    5: {"gfr": (5,  14), "creatinine": (5.0, 9.5), "weight": 0.25},
+}
+
+BLOOD_TYPES  = ["O+",  "A+",  "B+",  "AB+", "O-",  "A-",  "B-",  "AB-"]
+BLOOD_WGTS   = [0.38, 0.34,  0.09,  0.03,  0.07,  0.06,  0.02,  0.01]
+
+FIRST_NAMES = ["James","Maria","David","Sarah","Michael","Linda","Robert","Patricia",
+               "William","Barbara","Richard","Susan","Joseph","Jessica","Thomas","Karen",
+               "Charles","Lisa","Christopher","Nancy","Daniel","Betty","Matthew","Margaret",
+               "Anthony","Sandra","Mark","Ashley","Donald","Dorothy","Steven","Kimberly",
+               "Paul","Emily","Andrew","Donna","Joshua","Michelle","Kenneth","Carol",
+               "Kevin","Amanda","Brian","Melissa","George","Deborah","Timothy","Stephanie",
+               "Ronald","Rebecca"]
+LAST_NAMES  = ["Smith","Johnson","Williams","Brown","Jones","Garcia","Miller","Davis",
+               "Rodriguez","Martinez","Hernandez","Lopez","Gonzalez","Wilson","Anderson",
+               "Thomas","Taylor","Moore","Jackson","Martin","Lee","Perez","Thompson",
+               "White","Harris","Sanchez","Clark","Ramirez","Lewis","Robinson","Walker",
+               "Young","Allen","King","Wright","Scott","Torres","Nguyen","Hill","Flores",
+               "Green","Adams","Nelson","Baker","Hall","Rivera","Campbell","Mitchell",
+               "Carter","Roberts"]
+
+# ── Anomaly patients — fixed set for benchmark reproducibility ─────────────────
+# These patients have an extra measurement close to filing that triggers >25% delta
+ANOMALY_PATIENT_INDICES = {7, 12, 19, 26, 33}   # 0-indexed within P004-P050
+
+
+def pick_ckd_stage() -> int:
+    stages = list(CKD_STAGES.keys())
+    weights = [CKD_STAGES[s]["weight"] for s in stages]
+    return rng.choices(stages, weights=weights)[0]
+
+
+def generate_patient(idx: int) -> dict:
+    """Return a patient dict in prana_env format."""
+    patient_id = f"P{idx:03d}"
+    stage = pick_ckd_stage()
+    cfg = CKD_STAGES[stage]
+
+    # T5 current values
+    gfr_t5 = round(rng.uniform(*cfg["gfr"]), 1)
+    creatinine_t5 = round(rng.uniform(*cfg["creatinine"]), 1)
+
+    diabetic = rng.random() < 0.60
+    hba1c_t5 = round(rng.uniform(6.5, 12.0), 1) if diabetic else None
+
+    # Missing field scenarios
+    # Non-diabetic patients: 85% chance HbA1c not measured
+    missing_hba1c = (not diabetic and rng.random() < 0.85) or (diabetic and rng.random() < 0.04)
+    missing_creatinine = rng.random() < 0.05
+    missing_blood_type = rng.random() < 0.03
+
+    blood_type = rng.choices(BLOOD_TYPES, weights=BLOOD_WGTS)[0] if not missing_blood_type else None
+    pra = round(rng.uniform(0, 80), 1)
+
+    # T1 values: disease was less advanced — GFR higher, creatinine lower
+    gfr_t1 = round(min(gfr_t5 * rng.uniform(1.15, 1.60), 60.0), 1)
+    creatinine_t1 = round(creatinine_t5 * rng.uniform(0.55, 0.85), 1)
+    hba1c_t1 = round(hba1c_t5 * rng.uniform(0.82, 0.96), 1) if hba1c_t5 is not None else None
+
+    # Distractor fields — present, queryable, not KARS-required
+    cholesterol = round(rng.uniform(140, 270), 1)
+    bmi         = round(rng.uniform(18.5, 40.0), 1)
+    albumin     = round(rng.uniform(2.0, 4.2), 2)   # Low in CKD
+    hemoglobin  = round(rng.uniform(7.5, 13.5), 1)  # Anemia common in CKD
+
+    t1_snapshot: dict = {
+        "gfr": gfr_t1,
+        "creatinine": creatinine_t1 if not missing_creatinine else None,
+        "blood_type": blood_type,
+        "pra": pra,
+        "recorded_at": T1_NOMINAL_DATE.isoformat(),
+    }
+    if hba1c_t1 is not None and not missing_hba1c:
+        t1_snapshot["hba1c"] = hba1c_t1
+
+    patient: dict = {
+        "patient_id": patient_id,
+        "name": f"{rng.choice(FIRST_NAMES)} {rng.choice(LAST_NAMES)}",
+        "age": rng.randint(28, 72),
+        "ckd_stage": stage,
+        "gfr": gfr_t5,
+        "creatinine": creatinine_t5 if not missing_creatinine else None,
+        "blood_type": blood_type,
+        "pra": pra,
+        # Distractor fields
+        "cholesterol": cholesterol,
+        "bmi": bmi,
+        "albumin": albumin,
+        "hemoglobin": hemoglobin,
+        "t1_snapshot": t1_snapshot,
+    }
+    if hba1c_t5 is not None and not missing_hba1c:
+        patient["hba1c"] = hba1c_t5
+
+    return patient
+
+
+def to_tau2_patient(p: dict, anomaly: bool = False) -> dict:
+    """Convert prana_env patient dict to tau2 LabResult format."""
+    pid = p["patient_id"]
+
+    def lab_history(t1_val, t5_val, anomaly_entry=None) -> list:
+        entries = []
+        if t1_val is not None:
+            entries.append({"value": t1_val, "recorded_at": T1_NOMINAL_DATE.isoformat()})
+        if anomaly_entry:
+            entries.append(anomaly_entry)
+        if t5_val is not None:
+            entries.append({"value": t5_val, "recorded_at": T5_MEAS_DATE.isoformat()})
+        return entries
+
+    snap = p.get("t1_snapshot", {})
+
+    # Anomaly: inject a second T5 measurement with >25% delta, 6 days before filing
+    gfr_anomaly = None
+    if anomaly and p.get("gfr") is not None:
+        anomaly_gfr = round(p["gfr"] * 0.55, 1)  # 45% drop — clearly anomalous
+        gfr_anomaly = {"value": anomaly_gfr, "recorded_at": "2026-03-01"}
+
+    tau2: dict = {
+        "patient_id": pid,
+        "name": p["name"],
+        "age": p["age"],
+        "blood_type": p.get("blood_type"),
+        "pra": p.get("pra"),
+        "gfr": lab_history(snap.get("gfr"), p.get("gfr"), gfr_anomaly),
+        "creatinine": lab_history(snap.get("creatinine"), p.get("creatinine")),
+        "hba1c": lab_history(snap.get("hba1c"), p.get("hba1c")),
+    }
+    return tau2
+
+
+# ── Main ──────────────────────────────────────────────────────────────────────
+
+def main():
+    prana_env_root = Path(__file__).parent.parent
+    tau2_root = prana_env_root.parent / "tau2-bench"
+
+    # Load existing P001-P003 as anchors
+    existing_prana = json.loads((prana_env_root / "data" / "patient_db.json").read_text())
+    existing_patients = existing_prana["patients"]  # P001, P002, P003
+
+    # Load existing tau2 db to preserve non-patient sections
+    tau2_db_path = tau2_root / "data" / "tau2" / "domains" / "prana" / "db.json"
+    existing_tau2 = json.loads(tau2_db_path.read_text())
+
+    # Generate P004-P050
+    new_prana_patients = {}
+    new_tau2_patients = {}
+
+    for idx in range(4, N_PATIENTS + 1):
+        p = generate_patient(idx)
+        pid = p["patient_id"]
+        is_anomaly = (idx - 4) in ANOMALY_PATIENT_INDICES
+
+        # Add distractor fields to existing P001-P003 if not present
+        new_prana_patients[pid] = p
+        new_tau2_patients[pid] = to_tau2_patient(p, anomaly=is_anomaly)
+
+    # Add distractor fields to existing P001-P003
+    distractor_defaults = {
+        "P001": {"cholesterol": 187.3, "bmi": 24.1, "albumin": 3.2, "hemoglobin": 10.8},
+        "P002": {"cholesterol": 214.6, "bmi": 27.3, "albumin": 2.8, "hemoglobin": 9.4},
+        "P003": {"cholesterol": 168.9, "bmi": 22.7, "albumin": 3.6, "hemoglobin": 11.2},
+    }
+    for pid, extras in distractor_defaults.items():
+        existing_patients[pid].update(extras)
+
+    # ── Write prana_env patient_db.json ───────────────────────────────────────
+    prana_out = {"patients": {**existing_patients, **new_prana_patients}}
+    out_path = prana_env_root / "data" / "patient_db.json"
+    out_path.write_text(json.dumps(prana_out, indent=2))
+    print(f"Wrote {len(prana_out['patients'])} patients → {out_path}")
+
+    # ── Write tau2 db.json ─────────────────────────────────────────────────────
+    all_tau2_patients = {**existing_tau2["patient_db"], **new_tau2_patients}
+    existing_tau2["patient_db"] = all_tau2_patients
+    tau2_db_path.write_text(json.dumps(existing_tau2, indent=2))
+    print(f"Wrote {len(all_tau2_patients)} patients → {tau2_db_path}")
+
+    # ── Summary ───────────────────────────────────────────────────────────────
+    stages = {}
+    missing_hba1c = missing_creatinine = missing_bt = anomaly_count = 0
+    for pid, p in prana_out["patients"].items():
+        s = p.get("ckd_stage", "?")
+        stages[s] = stages.get(s, 0) + 1
+        if p.get("hba1c") is None:
+            missing_hba1c += 1
+        if p.get("creatinine") is None:
+            missing_creatinine += 1
+        if p.get("blood_type") is None:
+            missing_bt += 1
+
+    print(f"\nSummary:")
+    print(f"  CKD stages: {dict(sorted(stages.items()))}")
+    print(f"  Missing HbA1c:    {missing_hba1c}/{N_PATIENTS}")
+    print(f"  Missing creatinine: {missing_creatinine}/{N_PATIENTS}")
+    print(f"  Missing blood_type: {missing_bt}/{N_PATIENTS}")
+    print(f"  Anomaly patients (tau2): {len(ANOMALY_PATIENT_INDICES)}")
+
+
+if __name__ == "__main__":
+    main()

data/patient_db.json

@@ -16,7 +16,11 @@
         "blood_type": "A+",
         "pra": 12,
         "recorded_at": "2025-11-07"
-      }
+      },
+      "cholesterol": 187.3,
+      "bmi": 24.1,
+      "albumin": 3.2,
+      "hemoglobin": 10.8
     },
     "P002": {
       "patient_id": "P002",
@@ -34,7 +38,11 @@
         "blood_type": "O-",
         "pra": 45,
         "recorded_at": "2025-11-07"
-      }
+      },
+      "cholesterol": 214.6,
+      "bmi": 27.3,
+      "albumin": 2.8,
+      "hemoglobin": 9.4
     },
     "P003": {
       "patient_id": "P003",
@@ -52,7 +60,1058 @@
         "blood_type": "B+",
         "pra": 8,
         "recorded_at": "2025-11-07"
+      },
+      "cholesterol": 168.9,
+      "bmi": 22.7,
+      "albumin": 3.6,
+      "hemoglobin": 11.2
+    },
+    "P004": {
+      "patient_id": "P004",
+      "name": "Michelle Moore",
+      "age": 28,
+      "ckd_stage": 4,
+      "gfr": 15.4,
+      "creatinine": 3.2,
+      "blood_type": "A+",
+      "pra": 2.4,
+      "cholesterol": 165.8,
+      "bmi": 32.5,
+      "albumin": 3.2,
+      "hemoglobin": 8.8,
+      "t1_snapshot": {
+        "gfr": 19.2,
+        "creatinine": 2.2,
+        "blood_type": "A+",
+        "pra": 2.4,
+        "recorded_at": "2025-11-07",
+        "hba1c": 8.7
+      },
+      "hba1c": 10.6
+    },
+    "P005": {
+      "patient_id": "P005",
+      "name": "Kevin Flores",
+      "age": 51,
+      "ckd_stage": 5,
+      "gfr": 6.4,
+      "creatinine": 6.9,
+      "blood_type": "O+",
+      "pra": 27.5,
+      "cholesterol": 156.2,
+      "bmi": 38.3,
+      "albumin": 2.17,
+      "hemoglobin": 9.3,
+      "t1_snapshot": {
+        "gfr": 8.1,
+        "creatinine": 3.9,
+        "blood_type": "O+",
+        "pra": 27.5,
+        "recorded_at": "2025-11-07",
+        "hba1c": 6.8
+      },
+      "hba1c": 7.7
+    },
+    "P006": {
+      "patient_id": "P006",
+      "name": "Richard Wright",
+      "age": 43,
+      "ckd_stage": 4,
+      "gfr": 24.9,
+      "creatinine": 2.6,
+      "blood_type": "O+",
+      "pra": 50.9,
+      "cholesterol": 174.7,
+      "bmi": 38.6,
+      "albumin": 3.43,
+      "hemoglobin": 11.2,
+      "t1_snapshot": {
+        "gfr": 32.7,
+        "creatinine": 1.7,
+        "blood_type": "O+",
+        "pra": 50.9,
+        "recorded_at": "2025-11-07",
+        "hba1c": 6.9
+      },
+      "hba1c": 8.1
+    },
+    "P007": {
+      "patient_id": "P007",
+      "name": "Daniel Wilson",
+      "age": 69,
+      "ckd_stage": 3,
+      "gfr": 41.0,
+      "creatinine": 2.5,
+      "blood_type": "B+",
+      "pra": 18.3,
+      "cholesterol": 174.8,
+      "bmi": 23.0,
+      "albumin": 4.07,
+      "hemoglobin": 12.8,
+      "t1_snapshot": {
+        "gfr": 47.7,
+        "creatinine": 1.6,
+        "blood_type": "B+",
+        "pra": 18.3,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P008": {
+      "patient_id": "P008",
+      "name": "Richard Baker",
+      "age": 55,
+      "ckd_stage": 4,
+      "gfr": 27.4,
+      "creatinine": 4.1,
+      "blood_type": "A+",
+      "pra": 46.7,
+      "cholesterol": 204.2,
+      "bmi": 34.7,
+      "albumin": 3.89,
+      "hemoglobin": 8.4,
+      "t1_snapshot": {
+        "gfr": 36.0,
+        "creatinine": 3.5,
+        "blood_type": "A+",
+        "pra": 46.7,
+        "recorded_at": "2025-11-07",
+        "hba1c": 6.1
+      },
+      "hba1c": 7.3
+    },
+    "P009": {
+      "patient_id": "P009",
+      "name": "Timothy Taylor",
+      "age": 60,
+      "ckd_stage": 4,
+      "gfr": 20.4,
+      "creatinine": 4.0,
+      "blood_type": "O+",
+      "pra": 70.8,
+      "cholesterol": 178.2,
+      "bmi": 21.9,
+      "albumin": 2.01,
+      "hemoglobin": 11.8,
+      "t1_snapshot": {
+        "gfr": 30.4,
+        "creatinine": 3.1,
+        "blood_type": "O+",
+        "pra": 70.8,
+        "recorded_at": "2025-11-07",
+        "hba1c": 6.9
+      },
+      "hba1c": 7.9
+    },
+    "P010": {
+      "patient_id": "P010",
+      "name": "Margaret Martin",
+      "age": 43,
+      "ckd_stage": 5,
+      "gfr": 9.6,
+      "creatinine": 5.5,
+      "blood_type": "O+",
+      "pra": 12.9,
+      "cholesterol": 259.4,
+      "bmi": 31.4,
+      "albumin": 3.07,
+      "hemoglobin": 8.2,
+      "t1_snapshot": {
+        "gfr": 15.2,
+        "creatinine": 4.5,
+        "blood_type": "O+",
+        "pra": 12.9,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P011": {
+      "patient_id": "P011",
+      "name": "Joseph Rivera",
+      "age": 47,
+      "ckd_stage": 3,
+      "gfr": 55.5,
+      "creatinine": 2.4,
+      "blood_type": "O+",
+      "pra": 38.0,
+      "cholesterol": 195.0,
+      "bmi": 23.1,
+      "albumin": 3.19,
+      "hemoglobin": 11.9,
+      "t1_snapshot": {
+        "gfr": 60.0,
+        "creatinine": 1.5,
+        "blood_type": "O+",
+        "pra": 38.0,
+        "recorded_at": "2025-11-07",
+        "hba1c": 8.7
+      },
+      "hba1c": 9.2
+    },
+    "P012": {
+      "patient_id": "P012",
+      "name": "Betty Jones",
+      "age": 60,
+      "ckd_stage": 4,
+      "gfr": 24.4,
+      "creatinine": 3.4,
+      "blood_type": "O+",
+      "pra": 44.3,
+      "cholesterol": 232.0,
+      "bmi": 19.8,
+      "albumin": 2.15,
+      "hemoglobin": 7.7,
+      "t1_snapshot": {
+        "gfr": 34.5,
+        "creatinine": 1.9,
+        "blood_type": "O+",
+        "pra": 44.3,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P013": {
+      "patient_id": "P013",
+      "name": "Robert Brown",
+      "age": 53,
+      "ckd_stage": 3,
+      "gfr": 49.4,
+      "creatinine": 1.5,
+      "blood_type": "B+",
+      "pra": 15.2,
+      "cholesterol": 200.7,
+      "bmi": 34.2,
+      "albumin": 3.48,
+      "hemoglobin": 13.4,
+      "t1_snapshot": {
+        "gfr": 59.0,
+        "creatinine": 1.0,
+        "blood_type": "B+",
+        "pra": 15.2,
+        "recorded_at": "2025-11-07",
+        "hba1c": 10.2
+      },
+      "hba1c": 11.6
+    },
+    "P014": {
+      "patient_id": "P014",
+      "name": "Ronald Thomas",
+      "age": 38,
+      "ckd_stage": 4,
+      "gfr": 26.2,
+      "creatinine": 2.8,
+      "blood_type": "O-",
+      "pra": 6.0,
+      "cholesterol": 224.8,
+      "bmi": 30.1,
+      "albumin": 2.03,
+      "hemoglobin": 8.1,
+      "t1_snapshot": {
+        "gfr": 39.7,
+        "creatinine": 2.3,
+        "blood_type": "O-",
+        "pra": 6.0,
+        "recorded_at": "2025-11-07",
+        "hba1c": 7.8
+      },
+      "hba1c": 9.4
+    },
+    "P015": {
+      "patient_id": "P015",
+      "name": "Joseph Jackson",
+      "age": 41,
+      "ckd_stage": 4,
+      "gfr": 21.7,
+      "creatinine": null,
+      "blood_type": "A-",
+      "pra": 62.8,
+      "cholesterol": 235.0,
+      "bmi": 35.3,
+      "albumin": 3.46,
+      "hemoglobin": 10.4,
+      "t1_snapshot": {
+        "gfr": 27.7,
+        "creatinine": null,
+        "blood_type": "A-",
+        "pra": 62.8,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P016": {
+      "patient_id": "P016",
+      "name": "Patricia Torres",
+      "age": 43,
+      "ckd_stage": 5,
+      "gfr": 10.2,
+      "creatinine": 7.4,
+      "blood_type": "O-",
+      "pra": 12.6,
+      "cholesterol": 164.2,
+      "bmi": 31.3,
+      "albumin": 3.49,
+      "hemoglobin": 8.9,
+      "t1_snapshot": {
+        "gfr": 16.1,
+        "creatinine": 4.2,
+        "blood_type": "O-",
+        "pra": 12.6,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P017": {
+      "patient_id": "P017",
+      "name": "James Robinson",
+      "age": 67,
+      "ckd_stage": 4,
+      "gfr": 15.6,
+      "creatinine": 2.7,
+      "blood_type": "A+",
+      "pra": 25.1,
+      "cholesterol": 223.9,
+      "bmi": 28.3,
+      "albumin": 4.04,
+      "hemoglobin": 13.1,
+      "t1_snapshot": {
+        "gfr": 19.8,
+        "creatinine": 1.6,
+        "blood_type": "A+",
+        "pra": 25.1,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P018": {
+      "patient_id": "P018",
+      "name": "Nancy Nelson",
+      "age": 34,
+      "ckd_stage": 5,
+      "gfr": 14.0,
+      "creatinine": 5.3,
+      "blood_type": "O+",
+      "pra": 12.6,
+      "cholesterol": 268.3,
+      "bmi": 32.6,
+      "albumin": 2.02,
+      "hemoglobin": 12.4,
+      "t1_snapshot": {
+        "gfr": 21.4,
+        "creatinine": 4.0,
+        "blood_type": "O+",
+        "pra": 12.6,
+        "recorded_at": "2025-11-07",
+        "hba1c": 7.2
+      },
+      "hba1c": 8.0
+    },
+    "P019": {
+      "patient_id": "P019",
+      "name": "James Perez",
+      "age": 36,
+      "ckd_stage": 5,
+      "gfr": 6.2,
+      "creatinine": 5.5,
+      "blood_type": "O+",
+      "pra": 50.7,
+      "cholesterol": 250.0,
+      "bmi": 20.5,
+      "albumin": 2.93,
+      "hemoglobin": 9.2,
+      "t1_snapshot": {
+        "gfr": 7.9,
+        "creatinine": 3.8,
+        "blood_type": "O+",
+        "pra": 50.7,
+        "recorded_at": "2025-11-07",
+        "hba1c": 9.0
+      },
+      "hba1c": 9.5
+    },
+    "P020": {
+      "patient_id": "P020",
+      "name": "David Thompson",
+      "age": 41,
+      "ckd_stage": 4,
+      "gfr": 18.7,
+      "creatinine": 4.4,
+      "blood_type": "O-",
+      "pra": 43.6,
+      "cholesterol": 156.6,
+      "bmi": 25.1,
+      "albumin": 3.98,
+      "hemoglobin": 12.3,
+      "t1_snapshot": {
+        "gfr": 28.5,
+        "creatinine": 3.2,
+        "blood_type": "O-",
+        "pra": 43.6,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P021": {
+      "patient_id": "P021",
+      "name": "Brian Mitchell",
+      "age": 43,
+      "ckd_stage": 4,
+      "gfr": 24.3,
+      "creatinine": 3.4,
+      "blood_type": "O+",
+      "pra": 13.0,
+      "cholesterol": 163.3,
+      "bmi": 38.4,
+      "albumin": 3.72,
+      "hemoglobin": 10.0,
+      "t1_snapshot": {
+        "gfr": 36.7,
+        "creatinine": 2.1,
+        "blood_type": "O+",
+        "pra": 13.0,
+        "recorded_at": "2025-11-07",
+        "hba1c": 9.9
+      },
+      "hba1c": 11.3
+    },
+    "P022": {
+      "patient_id": "P022",
+      "name": "Rebecca Moore",
+      "age": 50,
+      "ckd_stage": 4,
+      "gfr": 26.0,
+      "creatinine": 2.8,
+      "blood_type": "A+",
+      "pra": 24.4,
+      "cholesterol": 143.1,
+      "bmi": 22.7,
+      "albumin": 2.72,
+      "hemoglobin": 12.7,
+      "t1_snapshot": {
+        "gfr": 39.2,
+        "creatinine": 1.7,
+        "blood_type": "A+",
+        "pra": 24.4,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P023": {
+      "patient_id": "P023",
+      "name": "Joshua Gonzalez",
+      "age": 44,
+      "ckd_stage": 4,
+      "gfr": 20.6,
+      "creatinine": 5.0,
+      "blood_type": "A-",
+      "pra": 21.2,
+      "cholesterol": 180.8,
+      "bmi": 31.5,
+      "albumin": 3.13,
+      "hemoglobin": 9.8,
+      "t1_snapshot": {
+        "gfr": 24.7,
+        "creatinine": 3.4,
+        "blood_type": "A-",
+        "pra": 21.2,
+        "recorded_at": "2025-11-07",
+        "hba1c": 10.8
+      },
+      "hba1c": 11.7
+    },
+    "P024": {
+      "patient_id": "P024",
+      "name": "Paul Martin",
+      "age": 70,
+      "ckd_stage": 3,
+      "gfr": 42.6,
+      "creatinine": 1.9,
+      "blood_type": "O+",
+      "pra": 34.5,
+      "cholesterol": 221.0,
+      "bmi": 32.8,
+      "albumin": 2.27,
+      "hemoglobin": 12.9,
+      "t1_snapshot": {
+        "gfr": 60.0,
+        "creatinine": 1.6,
+        "blood_type": "O+",
+        "pra": 34.5,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P025": {
+      "patient_id": "P025",
+      "name": "Donald Baker",
+      "age": 41,
+      "ckd_stage": 4,
+      "gfr": 20.6,
+      "creatinine": 3.2,
+      "blood_type": "A+",
+      "pra": 24.1,
+      "cholesterol": 195.9,
+      "bmi": 31.0,
+      "albumin": 3.44,
+      "hemoglobin": 10.3,
+      "t1_snapshot": {
+        "gfr": 28.8,
+        "creatinine": 1.8,
+        "blood_type": "A+",
+        "pra": 24.1,
+        "recorded_at": "2025-11-07",
+        "hba1c": 7.6
+      },
+      "hba1c": 8.8
+    },
+    "P026": {
+      "patient_id": "P026",
+      "name": "Steven Flores",
+      "age": 32,
+      "ckd_stage": 4,
+      "gfr": 26.1,
+      "creatinine": 4.9,
+      "blood_type": "A+",
+      "pra": 7.5,
+      "cholesterol": 180.4,
+      "bmi": 35.8,
+      "albumin": 2.32,
+      "hemoglobin": 7.8,
+      "t1_snapshot": {
+        "gfr": 41.2,
+        "creatinine": 3.0,
+        "blood_type": "A+",
+        "pra": 7.5,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P027": {
+      "patient_id": "P027",
+      "name": "Dorothy Brown",
+      "age": 63,
+      "ckd_stage": 4,
+      "gfr": 27.4,
+      "creatinine": 3.9,
+      "blood_type": "A+",
+      "pra": 71.4,
+      "cholesterol": 241.2,
+      "bmi": 23.2,
+      "albumin": 3.77,
+      "hemoglobin": 11.7,
+      "t1_snapshot": {
+        "gfr": 42.1,
+        "creatinine": 3.2,
+        "blood_type": "A+",
+        "pra": 71.4,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P028": {
+      "patient_id": "P028",
+      "name": "Stephanie Walker",
+      "age": 56,
+      "ckd_stage": 3,
+      "gfr": 54.6,
+      "creatinine": 1.8,
+      "blood_type": "A-",
+      "pra": 71.3,
+      "cholesterol": 205.6,
+      "bmi": 36.4,
+      "albumin": 3.21,
+      "hemoglobin": 12.9,
+      "t1_snapshot": {
+        "gfr": 60.0,
+        "creatinine": 1.4,
+        "blood_type": "A-",
+        "pra": 71.3,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P029": {
+      "patient_id": "P029",
+      "name": "Deborah Wilson",
+      "age": 32,
+      "ckd_stage": 4,
+      "gfr": 18.5,
+      "creatinine": 4.1,
+      "blood_type": "O+",
+      "pra": 22.9,
+      "cholesterol": 210.2,
+      "bmi": 21.5,
+      "albumin": 2.51,
+      "hemoglobin": 11.7,
+      "t1_snapshot": {
+        "gfr": 23.5,
+        "creatinine": 2.6,
+        "blood_type": "O+",
+        "pra": 22.9,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P030": {
+      "patient_id": "P030",
+      "name": "Andrew Mitchell",
+      "age": 62,
+      "ckd_stage": 4,
+      "gfr": 19.6,
+      "creatinine": 3.7,
+      "blood_type": "O+",
+      "pra": 70.5,
+      "cholesterol": 178.8,
+      "bmi": 26.9,
+      "albumin": 3.96,
+      "hemoglobin": 12.5,
+      "t1_snapshot": {
+        "gfr": 27.6,
+        "creatinine": 2.6,
+        "blood_type": "O+",
+        "pra": 70.5,
+        "recorded_at": "2025-11-07",
+        "hba1c": 10.6
+      },
+      "hba1c": 11.1
+    },
+    "P031": {
+      "patient_id": "P031",
+      "name": "Ashley Rodriguez",
+      "age": 57,
+      "ckd_stage": 5,
+      "gfr": 13.1,
+      "creatinine": 7.2,
+      "blood_type": "B+",
+      "pra": 67.2,
+      "cholesterol": 257.9,
+      "bmi": 31.2,
+      "albumin": 3.46,
+      "hemoglobin": 8.0,
+      "t1_snapshot": {
+        "gfr": 20.5,
+        "creatinine": 6.1,
+        "blood_type": "B+",
+        "pra": 67.2,
+        "recorded_at": "2025-11-07",
+        "hba1c": 8.2
+      },
+      "hba1c": 9.2
+    },
+    "P032": {
+      "patient_id": "P032",
+      "name": "David Carter",
+      "age": 29,
+      "ckd_stage": 3,
+      "gfr": 37.5,
+      "creatinine": 1.6,
+      "blood_type": "A+",
+      "pra": 66.8,
+      "cholesterol": 263.7,
+      "bmi": 23.3,
+      "albumin": 2.15,
+      "hemoglobin": 13.2,
+      "t1_snapshot": {
+        "gfr": 51.1,
+        "creatinine": 1.2,
+        "blood_type": "A+",
+        "pra": 66.8,
+        "recorded_at": "2025-11-07",
+        "hba1c": 6.9
+      },
+      "hba1c": 8.4
+    },
+    "P033": {
+      "patient_id": "P033",
+      "name": "Maria Martin",
+      "age": 64,
+      "ckd_stage": 5,
+      "gfr": 6.8,
+      "creatinine": 5.1,
+      "blood_type": "A+",
+      "pra": 61.4,
+      "cholesterol": 154.9,
+      "bmi": 36.1,
+      "albumin": 4.12,
+      "hemoglobin": 8.1,
+      "t1_snapshot": {
+        "gfr": 8.3,
+        "creatinine": 3.7,
+        "blood_type": "A+",
+        "pra": 61.4,
+        "recorded_at": "2025-11-07",
+        "hba1c": 6.7
+      },
+      "hba1c": 7.2
+    },
+    "P034": {
+      "patient_id": "P034",
+      "name": "Ashley Nelson",
+      "age": 51,
+      "ckd_stage": 4,
+      "gfr": 28.4,
+      "creatinine": 3.5,
+      "blood_type": "O+",
+      "pra": 61.8,
+      "cholesterol": 155.7,
+      "bmi": 39.7,
+      "albumin": 3.72,
+      "hemoglobin": 9.6,
+      "t1_snapshot": {
+        "gfr": 43.5,
+        "creatinine": 2.6,
+        "blood_type": "O+",
+        "pra": 61.8,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P035": {
+      "patient_id": "P035",
+      "name": "Timothy Nguyen",
+      "age": 45,
+      "ckd_stage": 3,
+      "gfr": 48.8,
+      "creatinine": 1.2,
+      "blood_type": "A+",
+      "pra": 12.2,
+      "cholesterol": 220.1,
+      "bmi": 38.3,
+      "albumin": 3.7,
+      "hemoglobin": 10.3,
+      "t1_snapshot": {
+        "gfr": 60.0,
+        "creatinine": 0.8,
+        "blood_type": "A+",
+        "pra": 12.2,
+        "recorded_at": "2025-11-07",
+        "hba1c": 8.4
+      },
+      "hba1c": 9.2
+    },
+    "P036": {
+      "patient_id": "P036",
+      "name": "Donna Smith",
+      "age": 61,
+      "ckd_stage": 4,
+      "gfr": 18.4,
+      "creatinine": 4.8,
+      "blood_type": "O+",
+      "pra": 39.5,
+      "cholesterol": 173.6,
+      "bmi": 24.5,
+      "albumin": 3.31,
+      "hemoglobin": 12.8,
+      "t1_snapshot": {
+        "gfr": 23.9,
+        "creatinine": 3.3,
+        "blood_type": "O+",
+        "pra": 39.5,
+        "recorded_at": "2025-11-07",
+        "hba1c": 7.8
+      },
+      "hba1c": 9.0
+    },
+    "P037": {
+      "patient_id": "P037",
+      "name": "Stephanie Scott",
+      "age": 49,
+      "ckd_stage": 5,
+      "gfr": 5.8,
+      "creatinine": 8.2,
+      "blood_type": "B+",
+      "pra": 7.4,
+      "cholesterol": 215.6,
+      "bmi": 28.7,
+      "albumin": 3.17,
+      "hemoglobin": 10.1,
+      "t1_snapshot": {
+        "gfr": 7.2,
+        "creatinine": 6.2,
+        "blood_type": "B+",
+        "pra": 7.4,
+        "recorded_at": "2025-11-07",
+        "hba1c": 9.2
+      },
+      "hba1c": 10.7
+    },
+    "P038": {
+      "patient_id": "P038",
+      "name": "Christopher Anderson",
+      "age": 51,
+      "ckd_stage": 4,
+      "gfr": 21.4,
+      "creatinine": 3.3,
+      "blood_type": "A+",
+      "pra": 15.3,
+      "cholesterol": 238.8,
+      "bmi": 31.3,
+      "albumin": 4.16,
+      "hemoglobin": 12.5,
+      "t1_snapshot": {
+        "gfr": 31.7,
+        "creatinine": 2.1,
+        "blood_type": "A+",
+        "pra": 15.3,
+        "recorded_at": "2025-11-07",
+        "hba1c": 9.5
+      },
+      "hba1c": 10.5
+    },
+    "P039": {
+      "patient_id": "P039",
+      "name": "Steven Davis",
+      "age": 32,
+      "ckd_stage": 3,
+      "gfr": 30.4,
+      "creatinine": 1.9,
+      "blood_type": "O-",
+      "pra": 39.3,
+      "cholesterol": 197.3,
+      "bmi": 22.5,
+      "albumin": 2.11,
+      "hemoglobin": 13.1,
+      "t1_snapshot": {
+        "gfr": 46.9,
+        "creatinine": 1.4,
+        "blood_type": "O-",
+        "pra": 39.3,
+        "recorded_at": "2025-11-07",
+        "hba1c": 10.5
+      },
+      "hba1c": 11.9
+    },
+    "P040": {
+      "patient_id": "P040",
+      "name": "Ashley Martin",
+      "age": 64,
+      "ckd_stage": 4,
+      "gfr": 16.0,
+      "creatinine": 4.1,
+      "blood_type": "O+",
+      "pra": 61.2,
+      "cholesterol": 207.9,
+      "bmi": 28.2,
+      "albumin": 2.97,
+      "hemoglobin": 12.7,
+      "t1_snapshot": {
+        "gfr": 22.8,
+        "creatinine": 3.2,
+        "blood_type": "O+",
+        "pra": 61.2,
+        "recorded_at": "2025-11-07",
+        "hba1c": 6.2
+      },
+      "hba1c": 7.3
+    },
+    "P041": {
+      "patient_id": "P041",
+      "name": "Deborah Jackson",
+      "age": 72,
+      "ckd_stage": 4,
+      "gfr": 23.5,
+      "creatinine": 4.4,
+      "blood_type": "O+",
+      "pra": 13.9,
+      "cholesterol": 198.6,
+      "bmi": 31.3,
+      "albumin": 2.64,
+      "hemoglobin": 8.9,
+      "t1_snapshot": {
+        "gfr": 27.8,
+        "creatinine": 2.4,
+        "blood_type": "O+",
+        "pra": 13.9,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P042": {
+      "patient_id": "P042",
+      "name": "Christopher Lewis",
+      "age": 35,
+      "ckd_stage": 5,
+      "gfr": 5.6,
+      "creatinine": 6.1,
+      "blood_type": "O+",
+      "pra": 18.0,
+      "cholesterol": 147.8,
+      "bmi": 35.5,
+      "albumin": 3.31,
+      "hemoglobin": 12.4,
+      "t1_snapshot": {
+        "gfr": 6.8,
+        "creatinine": 3.8,
+        "blood_type": "O+",
+        "pra": 18.0,
+        "recorded_at": "2025-11-07",
+        "hba1c": 9.2
+      },
+      "hba1c": 10.9
+    },
+    "P043": {
+      "patient_id": "P043",
+      "name": "Lisa Torres",
+      "age": 30,
+      "ckd_stage": 4,
+      "gfr": 19.3,
+      "creatinine": 3.5,
+      "blood_type": "B+",
+      "pra": 48.3,
+      "cholesterol": 251.9,
+      "bmi": 31.1,
+      "albumin": 2.05,
+      "hemoglobin": 12.1,
+      "t1_snapshot": {
+        "gfr": 22.4,
+        "creatinine": 2.2,
+        "blood_type": "B+",
+        "pra": 48.3,
+        "recorded_at": "2025-11-07",
+        "hba1c": 8.7
+      },
+      "hba1c": 9.5
+    },
+    "P044": {
+      "patient_id": "P044",
+      "name": "Sandra Carter",
+      "age": 63,
+      "ckd_stage": 5,
+      "gfr": 6.6,
+      "creatinine": 7.3,
+      "blood_type": "A-",
+      "pra": 37.6,
+      "cholesterol": 160.9,
+      "bmi": 27.4,
+      "albumin": 3.09,
+      "hemoglobin": 11.5,
+      "t1_snapshot": {
+        "gfr": 8.8,
+        "creatinine": 4.7,
+        "blood_type": "A-",
+        "pra": 37.6,
+        "recorded_at": "2025-11-07",
+        "hba1c": 6.7
+      },
+      "hba1c": 8.0
+    },
+    "P045": {
+      "patient_id": "P045",
+      "name": "Lisa Scott",
+      "age": 36,
+      "ckd_stage": 3,
+      "gfr": 32.6,
+      "creatinine": 1.5,
+      "blood_type": null,
+      "pra": 69.5,
+      "cholesterol": 220.9,
+      "bmi": 29.2,
+      "albumin": 2.97,
+      "hemoglobin": 7.8,
+      "t1_snapshot": {
+        "gfr": 44.3,
+        "creatinine": 0.8,
+        "blood_type": null,
+        "pra": 69.5,
+        "recorded_at": "2025-11-07",
+        "hba1c": 9.6
+      },
+      "hba1c": 10.7
+    },
+    "P046": {
+      "patient_id": "P046",
+      "name": "Deborah Jackson",
+      "age": 60,
+      "ckd_stage": 5,
+      "gfr": 8.9,
+      "creatinine": 8.1,
+      "blood_type": "A+",
+      "pra": 44.2,
+      "cholesterol": 200.0,
+      "bmi": 21.0,
+      "albumin": 3.83,
+      "hemoglobin": 10.5,
+      "t1_snapshot": {
+        "gfr": 10.6,
+        "creatinine": 6.9,
+        "blood_type": "A+",
+        "pra": 44.2,
+        "recorded_at": "2025-11-07",
+        "hba1c": 11.3
+      },
+      "hba1c": 11.9
+    },
+    "P047": {
+      "patient_id": "P047",
+      "name": "Timothy Martin",
+      "age": 65,
+      "ckd_stage": 4,
+      "gfr": 20.8,
+      "creatinine": 3.7,
+      "blood_type": "B+",
+      "pra": 10.9,
+      "cholesterol": 184.2,
+      "bmi": 35.4,
+      "albumin": 2.59,
+      "hemoglobin": 7.5,
+      "t1_snapshot": {
+        "gfr": 24.6,
+        "creatinine": 2.9,
+        "blood_type": "B+",
+        "pra": 10.9,
+        "recorded_at": "2025-11-07",
+        "hba1c": 8.5
+      },
+      "hba1c": 9.0
+    },
+    "P048": {
+      "patient_id": "P048",
+      "name": "Anthony Harris",
+      "age": 70,
+      "ckd_stage": 4,
+      "gfr": 24.2,
+      "creatinine": 4.7,
+      "blood_type": "A-",
+      "pra": 69.6,
+      "cholesterol": 251.3,
+      "bmi": 27.3,
+      "albumin": 2.7,
+      "hemoglobin": 10.3,
+      "t1_snapshot": {
+        "gfr": 38.5,
+        "creatinine": 2.9,
+        "blood_type": "A-",
+        "pra": 69.6,
+        "recorded_at": "2025-11-07",
+        "hba1c": 8.0
+      },
+      "hba1c": 9.2
+    },
+    "P049": {
+      "patient_id": "P049",
+      "name": "Michael Walker",
+      "age": 44,
+      "ckd_stage": 5,
+      "gfr": 10.9,
+      "creatinine": 5.7,
+      "blood_type": "O+",
+      "pra": 67.6,
+      "cholesterol": 142.0,
+      "bmi": 21.6,
+      "albumin": 3.92,
+      "hemoglobin": 13.3,
+      "t1_snapshot": {
+        "gfr": 13.0,
+        "creatinine": 4.7,
+        "blood_type": "O+",
+        "pra": 67.6,
+        "recorded_at": "2025-11-07"
+      }
+    },
+    "P050": {
+      "patient_id": "P050",
+      "name": "Thomas Mitchell",
+      "age": 33,
+      "ckd_stage": 4,
+      "gfr": 24.7,
+      "creatinine": 4.1,
+      "blood_type": "O+",
+      "pra": 57.5,
+      "cholesterol": 144.7,
+      "bmi": 20.0,
+      "albumin": 3.39,
+      "hemoglobin": 13.0,
+      "t1_snapshot": {
+        "gfr": 36.8,
+        "creatinine": 3.3,
+        "blood_type": "O+",
+        "pra": 57.5,
+        "recorded_at": "2025-11-07"
       }
     }
   }
-}
+}

prana_grpo_qwen3_8b_fp8.ipynb

@@ -0,0 +1,379 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# PRANA-Env: Reinforcement Learning with Qwen3-8B FP8\n",
+    "\n",
+    "Fine-tune **Qwen3-8B** using **GRPO + FP8** on the PRANA kidney transplant administration environment.\n",
+    "\n",
+    "The agent must:\n",
+    "1. Query fragmented clinical datastores\n",
+    "2. Detect stale lab values (90-day KARS recency window)\n",
+    "3. Detect anomalous measurements (>25% change within 14 days)\n",
+    "4. File a complete KARS-compliant SRTR report\n",
+    "\n",
+    "Reward signal comes from the deterministic KARS validator in prana_env.\n",
+    "\n",
+    "**Hardware**: H100 required for FP8.\n",
+    "\n",
+    "**Baseline**: Qwen3-8B untuned scores **0.71 Pass@1** on temporal/anomaly tasks.  \n",
+    "**Target**: ≥ 0.90 Pass@1 after GRPO fine-tuning."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 1. Installation"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": "import os\nos.environ['UNSLOTH_VLLM_STANDBY'] = '1'\n\nfrom unsloth import FastLanguageModel\nimport torch\n\nmax_seq_length = 2048  # Multi-turn clinical dialogue\nlora_rank = 32         # From official Qwen3-8B FP8 notebook\n\nmodel, tokenizer = FastLanguageModel.from_pretrained(\n    model_name = 'unsloth/Qwen3-8B-FP8',\n    max_seq_length = max_seq_length,\n    load_in_4bit = False,       # FP8, not 4bit\n    fast_inference = True,      # vLLM fast inference for GRPO rollouts\n    max_lora_rank = lora_rank,\n    load_in_fp8 = True,         # FP8 training on H100\n)\n\nmodel = FastLanguageModel.get_peft_model(\n    model,\n    r = lora_rank,\n    target_modules = [\n        'q_proj', 'k_proj', 'v_proj', 'o_proj',\n        'gate_proj', 'up_proj', 'down_proj',\n    ],\n    lora_alpha = lora_rank * 2,\n    use_gradient_checkpointing = 'unsloth',\n    random_state = 3407,\n)"
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%capture\n",
+    "# Clone prana_env and install dependencies\n",
+    "!git clone https://github.com/pbanavara/prana_env.git\n",
+    "!pip install -q fastapi uvicorn websockets pydantic openenv requests\n",
+    "%cd prana_env\n",
+    "!pip install -q -e .\n",
+    "\n",
+    "import sys, os\n",
+    "sys.path.insert(0, '.')\n",
+    "working_directory = os.path.abspath('.')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 2. Load Qwen3-8B with FP8 + LoRA"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import os\n",
+    "os.environ['UNSLOTH_VLLM_STANDBY'] = '1'\n",
+    "\n",
+    "from unsloth import FastLanguageModel\n",
+    "import torch\n",
+    "\n",
+    "max_seq_length = 2048  # Multi-turn clinical dialogue\n",
+    "lora_rank = 32         # From official Qwen3-8B FP8 notebook\n",
+    "\n",
+    "model, tokenizer = FastLanguageModel.from_pretrained(\n",
+    "    model_name = 'unsloth/Qwen3-8B',\n",
+    "    max_seq_length = max_seq_length,\n",
+    "    load_in_4bit = False,       # FP8, not 4bit\n",
+    "    fast_inference = True,      # vLLM fast inference for GRPO rollouts\n",
+    "    max_lora_rank = lora_rank,\n",
+    "    load_in_fp8 = True,         # FP8 training on H100\n",
+    ")\n",
+    "\n",
+    "model = FastLanguageModel.get_peft_model(\n",
+    "    model,\n",
+    "    r = lora_rank,\n",
+    "    target_modules = [\n",
+    "        'q_proj', 'k_proj', 'v_proj', 'o_proj',\n",
+    "        'gate_proj', 'up_proj', 'down_proj',\n",
+    "    ],\n",
+    "    lora_alpha = lora_rank * 2,\n",
+    "    use_gradient_checkpointing = 'unsloth',\n",
+    "    random_state = 3407,\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 3. Launch prana_env server\n",
+    "\n",
+    "Start the FastAPI + WebSocket server as a local subprocess."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": "import subprocess, time, requests\n\nPRANA_PORT = 7860\nPRANA_BASE_URL = f'http://localhost:{PRANA_PORT}'\n_server_proc = None\n\ndef launch_prana_server():\n    global _server_proc\n    if _server_proc is not None:\n        try:\n            requests.get(f'{PRANA_BASE_URL}/health', timeout=2)\n            return\n        except Exception:\n            _server_proc.kill()\n            _server_proc = None\n\n    _server_proc = subprocess.Popen(\n        ['uvicorn', 'server.app:app', '--host', '0.0.0.0', '--port', str(PRANA_PORT)],\n        cwd=working_directory,\n        stdout=subprocess.DEVNULL,\n        stderr=subprocess.DEVNULL,\n    )\n    for _ in range(20):\n        try:\n            requests.get(f'{PRANA_BASE_URL}/health', timeout=2)\n            print(f'prana_env server ready at {PRANA_BASE_URL}')\n            return\n        except Exception:\n            time.sleep(1)\n    raise RuntimeError('prana_env server failed to start')\n\nlaunch_prana_server()"
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 4. PRANA-Env client helpers"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": "import random\nfrom prana_env.client import PranaEnv\nfrom prana_env.models import PranaAction\n\nPATIENTS = [f'P{i:03d}' for i in range(1, 51)]  # P001-P050\n\ndef run_episode(action_sequence: list[dict], patient_id: str) -> tuple[float, str]:\n    \"\"\"\n    Execute a list of parsed actions against prana_env.\n    Returns (cumulative_reward, 'PASSED'|'FAILED'|'INCOMPLETE').\n    \"\"\"\n    launch_prana_server()\n    cumulative_reward = 0.0\n    kars_result = 'INCOMPLETE'\n\n    with PranaEnv(base_url=PRANA_BASE_URL) as env:\n        env.reset(patient_id=patient_id)\n        for action_dict in action_sequence:\n            try:\n                action = PranaAction(**action_dict)\n                result = env.step(action)\n                cumulative_reward += result.reward\n                if result.done:\n                    kars_result = result.observation.kars_result or 'FAILED'\n                    break\n            except Exception:\n                cumulative_reward -= 1.0\n                continue\n\n    return cumulative_reward, kars_result"
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 5. Action parser"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import json, re\n",
+    "\n",
+    "def extract_actions(response: str) -> list[dict]:\n",
+    "    \"\"\"\n",
+    "    Extract a JSON array of actions from the model response.\n",
+    "    Model is instructed to output actions inside ```json ... ``` blocks.\n",
+    "    \"\"\"\n",
+    "    match = re.search(r'```json\\s*(\\[.*?\\])\\s*```', response, re.DOTALL)\n",
+    "    if not match:\n",
+    "        match = re.search(r'(\\[\\s*\\{.*?\\}\\s*\\])', response, re.DOTALL)\n",
+    "    if not match:\n",
+    "        return []\n",
+    "    try:\n",
+    "        return json.loads(match.group(1))\n",
+    "    except json.JSONDecodeError:\n",
+    "        return []"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 6. GRPO prompt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "SYSTEM_PROMPT = \"\"\"\n",
+    "You are a clinical administrative agent for a kidney transplant center.\n",
+    "Your task is to file a KARS-compliant SRTR report for a patient.\n",
+    "\n",
+    "Today's date is 2026-03-07 (filing date T5).\n",
+    "The patient has a record from approximately 4 months ago (T1). Some fields may be stale.\n",
+    "\n",
+    "KARS Required Fields:\n",
+    "- hba1c, gfr, creatinine (PatientDB) — time-sensitive, must be within 90 days of filing\n",
+    "- blood_type (PatientDB) — stable, no recency requirement\n",
+    "\n",
+    "OPTN Clinical Integrity Policy:\n",
+    "- If two measurements of the same field within 14 days differ by >25%, do NOT file.\n",
+    "  Communicate the anomaly and recommend a confirmatory test.\n",
+    "\n",
+    "Actions available:\n",
+    "- query_db: {action_type: query_db, target: PatientDB, field: <field>, patient_id: <id>}\n",
+    "- record_value: {action_type: record_value, field: <field>, value: <value>}\n",
+    "- file_report: {action_type: file_report}\n",
+    "\n",
+    "Output your complete action plan as a JSON array inside ```json ... ``` tags.\n",
+    "Reason step by step before outputting actions.\n",
+    "\"\"\".strip()\n",
+    "\n",
+    "USER_PROMPT_TEMPLATE = \"\"\"\n",
+    "File a KARS-compliant SRTR report for patient {patient_id}.\n",
+    "The T1 snapshot from ~4 months ago is pre-loaded in the record.\n",
+    "Check which fields are stale or anomalous, re-query only what is needed, and file.\n",
+    "\"\"\".strip()\n",
+    "\n",
+    "def make_prompt(patient_id: str) -> list[dict]:\n",
+    "    return [\n",
+    "        {'role': 'system', 'content': SYSTEM_PROMPT},\n",
+    "        {'role': 'user', 'content': USER_PROMPT_TEMPLATE.format(patient_id=patient_id)},\n",
+    "    ]"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 7. Reward functions"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": "def actions_parseable(completions, **kwargs):\n    \"\"\"Reward 1.0 if model outputs a parseable action list, -1.0 otherwise.\"\"\"\n    scores = []\n    for completion in completions:\n        response = completion[0]['content']\n        actions = extract_actions(response)\n        scores.append(1.0 if len(actions) > 0 else -1.0)\n    return scores\n\n\ndef kars_reward(completions, prompts, **kwargs):\n    \"\"\"\n    Execute the action sequence in prana_env and return the KARS reward.\n    prana_env reward scale:\n      +15  KARS PASSED first attempt\n      +10  KARS PASSED after correction\n       -1  re-query of already-fresh field\n       -5  KARS FAILED\n      -10  unrecoverable (3 attempts)\n    Normalized to [-1, 1] for GRPO stability.\n    \"\"\"\n    scores = []\n    for completion, prompt in zip(completions, prompts):\n        response = completion[0]['content']\n        actions = extract_actions(response)\n\n        if not actions:\n            scores.append(-1.0)\n            continue\n\n        # Extract patient_id from user message (P001-P050)\n        patient_id = 'P001'\n        for msg in prompt:\n            if msg['role'] == 'user':\n                m = re.search(r'P\\d{3}', msg['content'])\n                if m:\n                    patient_id = m.group(0)\n\n        # Inject patient_id into query_db actions if missing\n        for a in actions:\n            if a.get('action_type') == 'query_db' and 'patient_id' not in a:\n                a['patient_id'] = patient_id\n\n        try:\n            raw_reward, kars_result = run_episode(actions, patient_id)\n            normalized = max(-1.0, min(1.0, raw_reward / 15.0))\n            scores.append(normalized)\n            print(f'[KARS] patient={patient_id} result={kars_result} raw={raw_reward:.1f} norm={normalized:.2f}')\n        except Exception as e:\n            print(f'[KARS] error: {e}')\n            scores.append(-1.0)\n\n    return scores"
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 8. Dataset"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": "from datasets import Dataset\n\n# 1000 episodes cycling through all 50 patients\nrecords = []\nfor i in range(1000):\n    pid = PATIENTS[i % len(PATIENTS)]\n    records.append({\n        'prompt': make_prompt(pid),\n        'answer': 0,\n    })\n\ndataset = Dataset.from_list(records)\n\nmaximum_length = len(tokenizer.apply_chat_template(\n    make_prompt('P001'),\n    add_generation_prompt=True,\n))\nprint(f'Prompt token length: {maximum_length}')\ndataset[0]"
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 9. GRPO Training"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "max_prompt_length = maximum_length + 1\n",
+    "max_completion_length = max_seq_length - max_prompt_length\n",
+    "\n",
+    "from vllm import SamplingParams\n",
+    "vllm_sampling_params = SamplingParams(\n",
+    "    min_p = 0.1,\n",
+    "    top_p = 1.0,\n",
+    "    top_k = -1,\n",
+    "    seed = 3407,\n",
+    "    stop = [tokenizer.eos_token],\n",
+    "    include_stop_str_in_output = True,\n",
+    ")\n",
+    "\n",
+    "from trl import GRPOConfig, GRPOTrainer\n",
+    "\n",
+    "training_args = GRPOConfig(\n",
+    "    vllm_sampling_params = vllm_sampling_params,\n",
+    "    temperature = 1.0,\n",
+    "    learning_rate = 5e-6,\n",
+    "    weight_decay = 0.01,\n",
+    "    warmup_ratio = 0.1,\n",
+    "    lr_scheduler_type = 'linear',\n",
+    "    optim = 'adamw_8bit',\n",
+    "    logging_steps = 1,\n",
+    "    per_device_train_batch_size = 4,\n",
+    "    gradient_accumulation_steps = 1,\n",
+    "    num_generations = 4,        # Increase to 8 if VRAM allows\n",
+    "    max_prompt_length = max_prompt_length,\n",
+    "    max_completion_length = max_completion_length,\n",
+    "    max_steps = 600,\n",
+    "    save_steps = 100,\n",
+    "    report_to = 'none',\n",
+    "    output_dir = 'outputs',\n",
+    ")\n",
+    "\n",
+    "trainer = GRPOTrainer(\n",
+    "    model = model,\n",
+    "    processing_class = tokenizer,\n",
+    "    reward_funcs = [\n",
+    "        actions_parseable,\n",
+    "        kars_reward,\n",
+    "    ],\n",
+    "    args = training_args,\n",
+    "    train_dataset = dataset,\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "trainer.train()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 10. Inference — test the fine-tuned model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from transformers import TextStreamer\n",
+    "\n",
+    "test_patient = 'P002'  # anomalous GFR/creatinine — hardest case\n",
+    "text = tokenizer.apply_chat_template(\n",
+    "    make_prompt(test_patient),\n",
+    "    tokenize=False,\n",
+    "    add_generation_prompt=True,\n",
+    ")\n",
+    "\n",
+    "_ = model.generate(\n",
+    "    **tokenizer(text, return_tensors='pt').to('cuda'),\n",
+    "    temperature=1.0,\n",
+    "    max_new_tokens=1024,\n",
+    "    streamer=TextStreamer(tokenizer, skip_prompt=False),\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 11. Save model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "model.save_pretrained('prana_qwen3_8b_lora')\n",
+    "tokenizer.save_pretrained('prana_qwen3_8b_lora')\n",
+    "\n",
+    "# Push to Hub (optional)\n",
+    "if False:\n",
+    "    model.push_to_hub_merged(\n",
+    "        'pbanavara/prana-qwen3-8b-grpo',\n",
+    "        tokenizer,\n",
+    "        save_method='merged_16bit',\n",
+    "        token='hf_...',\n",
+    "    )"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "name": "python",
+   "version": "3.11.0"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}

server/app.py

@@ -19,10 +19,10 @@ Endpoints:
 
 Usage:
     # Development (with auto-reload):
-    uvicorn server.app:app --reload --host 0.0.0.0 --port 8000
+    uvicorn server.app:app --reload --host 0.0.0.0 --port 7860
 
     # Production:
-    uvicorn server.app:app --host 0.0.0.0 --port 8000 --workers 4
+    uvicorn server.app:app --host 0.0.0.0 --port 7860 --workers 4
 
     # Or run directly:
     python -m server.app
@@ -50,7 +50,7 @@ app = create_app(
 )
 
 
-def main(host: str = "0.0.0.0", port: int = 8000):
+def main(host: str = "0.0.0.0", port: int = 7860):
     """
     Entry point for direct execution via uv run or python -m.
 
@@ -76,6 +76,6 @@ if __name__ == "__main__":
     import argparse
 
     parser = argparse.ArgumentParser()
-    parser.add_argument("--port", type=int, default=8000)
+    parser.add_argument("--port", type=int, default=7860)
     args = parser.parse_args()
     main(port=args.port)

server/prana_env_environment.py

@@ -327,35 +327,19 @@ class PranaEnvironment(Environment):
             history.append((anom_date, self._injected_anomaly["value"]))
 
         history.append((EPISODE_DATE, float(t5_value)))
-        history.sort(key=lambda x: x[0])
 
-        lines = []
-        for i, (d, v) in enumerate(history):
-            suffix = " ← latest" if i == len(history) - 1 else ""
-            lines.append(f"  {v} (recorded: {d}){suffix}")
+        # Shuffle deterministically by (patient_id, field) — agent must sort by date.
+        # No ← latest pointer, no anomaly flag — matches tau2 benchmark behaviour.
+        rng = random.Random(hash((patient_id, field)) & 0xFFFFFFFF)
+        rng.shuffle(history)
 
-        result = (
+        lines = [f"  {v} (recorded: {d})" for d, v in history]
+
+        return (
             f"{field} measurement history for {patient_id} "
             f"(filing date: {EPISODE_DATE}):\n" + "\n".join(lines)
         )
 
-        # Check for anomaly between consecutive entries within window
-        for i in range(len(history) - 1):
-            d1, v1 = history[i]
-            d2, v2 = history[i + 1]
-            days_apart = (d2 - d1).days
-            if days_apart <= ANOMALY_WINDOW_DAYS and v1 > 0:
-                change = abs(v2 - v1) / v1
-                if change >= ANOMALY_THRESHOLD:
-                    pct = round(change * 100, 1)
-                    result += (
-                        f"\n⚠️ ANOMALY DETECTED: {v1} ({d1}) → {v2} ({d2}), "
-                        f"{days_apart} days apart, {pct}% delta. "
-                        f"Recommend confirmatory test before filing."
-                    )
-
-        return result
-
     def _handle_record_value(self, action: PranaAction) -> PranaObservation:
         field = (action.field or "").lower()
         value = action.value