Create scorer.py

scorer.py

@@ -0,0 +1,69 @@
+# scorer.py
+import csv, json, re, sys
+
+ALLOWED = {"A","B"}
+
+def norm(t):
+    return re.sub(r"\s+"," ",(t or "").lower().strip())
+
+def kws(s):
+    return [k.strip().lower() for k in (s or "").split("|") if k.strip()]
+
+def load_gold(p):
+    g={}
+    with open(p, newline="", encoding="utf-8") as f:
+        r=csv.DictReader(f)
+        for d in r:
+            g[d["sample_id"]] = {
+                "opt": d["correct_option"].strip().upper(),
+                "kws": kws(d.get("required_keywords","")),
+                "axis": d.get("axis","unknown") or "unknown"
+            }
+    return g
+
+def load_pred(p):
+    o={}
+    with open(p, newline="", encoding="utf-8") as f:
+        r=csv.DictReader(f)
+        for d in r:
+            o[d["sample_id"]] = {
+                "opt": d.get("predicted_option","").strip().upper(),
+                "note": d.get("predicted_note","")
+            }
+    return o
+
+def cov(note, ks):
+    if not ks: return 1.0
+    n=norm(note)
+    return sum(1 for k in ks if k in n)/len(ks)
+
+def main():
+    g=load_gold(sys.argv[1])
+    p=load_pred(sys.argv[2])
+
+    tot=len(g)
+    ok=0; miss=0; bad=0
+    covs=[]
+
+    for sid,gd in g.items():
+        pr=p.get(sid)
+        if not pr:
+            miss+=1; covs.append(0); continue
+        if pr["opt"] not in ALLOWED:
+            bad+=1
+        elif pr["opt"]==gd["opt"]:
+            ok+=1
+        covs.append(cov(pr["note"], gd["kws"]))
+
+    out={
+        "option_accuracy": ok/max(1,tot),
+        "note_mean_keyword_coverage": sum(covs)/max(1,tot),
+        "joint_success_rate": (ok/max(1,tot))*(sum(1 for c in covs if c>=0.6)/max(1,tot)),
+        "total": tot,
+        "missing": miss,
+        "invalid": bad
+    }
+    print(json.dumps(out, indent=2))
+
+if __name__=="__main__":
+    main()