Bugfix: WebSRC should be token-level F1 NOT character-level

lmms_eval/tasks/websrc/utils.py

@@ -50,7 +50,7 @@ def websrc_process_results(doc, results):
         "websrc_squad_f1": websrc_ans,
         "submission": {
             websrc_ans['question_id']: pred,
-        },
+        } if 'question_id' in websrc_ans else None
     }
 
 
@@ -122,27 +122,39 @@ def _normalize_str(string):
         # lower it
         string = string.lower()
 
-        # strip non-alphanumeric characters
-        string = re.sub(r"[^a-zA-Z0-9]", "", string)
-
         # strip leading and trailing whitespaces
         string = string.strip()
 
         return string
 
+    def _tokenize(text):
+        # Regex pattern to match words and isolate punctuation
+        pattern = r'\w+|[^\w\s]'
+        tokens = re.findall(pattern, text)
+        return tokens
+
+    def _compute_f1(sa, sb):
+        sa = _normalize_str(sa)
+        sb = _normalize_str(sb)
+
+        sa = _tokenize(sa)
+        sb = _tokenize(sb)
+
+        sa = set(sa)
+        sb = set(sb)
+
+        if len(sa) == 0 or len(sb) == 0:
+            return 0.0
+
+        comm = sa.intersection(sb)
+        prec = len(comm) / len(sb)
+        rec = len(comm) / len(sa)
+        f1 = 2 * prec * rec / (prec + rec) if prec + rec > 0 else 0
+        return f1
+
     judge_list = []
     for sample in samples:
-        gold_i = set(_normalize_str(sample["answer"]))
-        pred_i = set(_normalize_str( sample["parsed_pred"]))
-        if len(pred_i) == 0:
-            judge_list.append(0.0)
-            continue
-
-        comm_i = gold_i.intersection(pred_i)
-        prec_i = len(comm_i) / len(pred_i)
-        rec_i = len(comm_i) / len(gold_i)
-        f1_i = 2 * prec_i * rec_i / (prec_i + rec_i) if prec_i + rec_i > 0 else 0
-        judge_list.append(f1_i)
+        judge_list.append(_compute_f1(sample["answer"], sample["parsed_pred"]))
 
     f1 = np.mean(judge_list)
     return judge_list, {"f1": f1}