Create scorer.py

scorer.py

@@ -0,0 +1,116 @@
+import csv
+import json
+import re
+from typing import Dict, Tuple
+
+ALLOWED_TYPES = {
+    "low_uncertainty",
+    "moderate_confidence",
+    "confidence_borderline",
+    "confidence_low",
+    "disorder_high",
+    "model_disagreement",
+    "high_disorder",
+}
+
+def norm(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+def token_set(s: str) -> set:
+    s = norm(s)
+    s = re.sub(r"[^a-z0-9\s]", " ", s)
+    return set([t for t in s.split(" ") if t])
+
+def jaccard(a: str, b: str) -> float:
+    ta, tb = token_set(a), token_set(b)
+    if not ta or not tb:
+        return 0.0
+    return len(ta & tb) / len(ta | tb)
+
+def load_refs(test_csv):
+    refs={}
+    with open(test_csv,encoding="utf-8") as f:
+        r=csv.DictReader(f)
+        for row in r:
+            refs[row["id"]] = (
+                norm(row["gold_uncertainty_flag"]),
+                row["gold_uncertainty_type"],
+                row["gold_recommendation"],
+            )
+    return refs
+
+def extract_json(text):
+    try:
+        return json.loads(text)
+    except:
+        m=re.search(r"\{.*\}",text)
+        if m:
+            try:
+                return json.loads(m.group(0))
+            except:
+                return {}
+    return {}
+
+def score(pred_path,test_csv):
+    refs=load_refs(test_csv)
+
+    n=0
+    flag_hits=0
+    type_hits=0
+    rec_sim=0
+    fmt_hits=0
+
+    with open(pred_path,encoding="utf-8") as f:
+        preds=[json.loads(x) for x in f if x.strip()]
+
+    for p in preds:
+        pid=p["id"]
+        if pid not in refs:
+            continue
+
+        n+=1
+        parsed=extract_json(p["prediction"])
+
+        pred_flag=norm(parsed.get("uncertainty_flag"))
+        pred_type=parsed.get("uncertainty_type","")
+        pred_rec=parsed.get("recommendation","")
+
+        gold_flag,gold_type,gold_rec=refs[pid]
+
+        if pred_flag==gold_flag:
+            flag_hits+=1
+
+        if pred_type==gold_type:
+            type_hits+=1
+
+        rec_sim+=jaccard(pred_rec,gold_rec)
+
+        if pred_flag in {"yes","no"} and pred_type in ALLOWED_TYPES:
+            fmt_hits+=1
+
+    if n==0:
+        return {"final_score":0}
+
+    acc_flag=flag_hits/n
+    acc_type=type_hits/n
+    sim_rec=rec_sim/n
+    fmt=fmt_hits/n
+
+    final=0.4*acc_flag+0.3*acc_type+0.2*sim_rec+0.1*fmt
+
+    return {
+        "final_score":final,
+        "uncertainty_flag_accuracy":acc_flag,
+        "uncertainty_type_accuracy":acc_type,
+        "recommendation_similarity":sim_rec,
+        "format_pass_rate":fmt,
+        "n":n
+    }
+
+if __name__=="__main__":
+    import argparse
+    p=argparse.ArgumentParser()
+    p.add_argument("--predictions")
+    p.add_argument("--test_csv")
+    args=p.parse_args()
+    print(score(args.predictions,args.test_csv))