Smoke job repo: grouped tabular dataset + config

400 rows / 200 patients, two samples each, so grouped splitting is actually exercised. Signal is modest and noisy on purpose — a separable toy would score 1.0 with a broken model and prove nothing.
2026-08-13 21:16:22 +02:00
commit 4e815df327
5 changed files with 603 additions and 0 deletions
@@ -0,0 +1,51 @@
+"""Generate the smoke dataset. Deterministic — rerunning reproduces the file byte
+for byte, so the committed CSV can always be regenerated and diffed.
+
+Signal is real but modest: the label depends on age/iop/cdr with noise, so a
+working pipeline lands around 0.85-0.95 AUC and a broken one sits near 0.5. A
+perfectly separable toy would hide bugs by scoring 1.0 no matter what.
+"""
+import csv, math, random
+from pathlib import Path
+
+rng = random.Random(20260813)
+N_PATIENTS, EYES = 200, 2
+SITES = ["skovde", "gothenburg", "linkoping"]
+out = Path(__file__).resolve().parents[1] / "data" / "smoke" / "labels.csv"
+
+rows = []
+for p in range(N_PATIENTS):
+    # Patient-level traits: both samples from one patient share them, which is
+    # what makes grouped splitting matter — a per-row split would leak.
+    base_age = rng.gauss(62, 11)
+    site = rng.choice(SITES)
+    frailty = rng.gauss(0, 1)
+    for e in range(EYES):
+        age = base_age + rng.gauss(0, 1.5)
+        iop = rng.gauss(16 + 2.2 * frailty, 2.8)        # intraocular pressure
+        cdr = min(0.95, max(0.05, rng.gauss(0.45 + 0.09 * frailty, 0.11)))  # cup/disc
+        rnfl = rng.gauss(95 - 7 * frailty, 9)           # retinal nerve fibre layer
+        noise = rng.gauss(0, 1)
+        logit = -6.0 + 0.045 * age + 0.20 * iop + 4.2 * cdr - 0.035 * rnfl + 0.5 * noise
+        pos = rng.random() < 1 / (1 + math.exp(-logit))
+        rows.append({
+            "sample_id":  f"S{p:03d}_{e}",
+            "patient_id": f"P{p:03d}",
+            "age":        round(age, 1),
+            "iop":        round(iop, 1),
+            "cdr":        round(cdr, 3),
+            "rnfl":       round(rnfl, 1),
+            "noise_feat": round(noise, 3),
+            "site":       site,
+            "notes":      "routine screening" if not pos else "referred",
+            "diagnosis":  "glaucoma" if pos else "healthy",
+        })
+
+out.parent.mkdir(parents=True, exist_ok=True)
+with out.open("w", newline="") as fh:
+    w = csv.DictWriter(fh, fieldnames=list(rows[0]))
+    w.writeheader()
+    w.writerows(rows)
+
+pos = sum(r["diagnosis"] == "glaucoma" for r in rows)
+print(f"wrote {out}  rows={len(rows)}  patients={N_PATIENTS}  positives={pos} ({pos/len(rows):.0%})")