round average

baseline_scores.json

@@ -4,5 +4,5 @@
     "task2": 0.99,
     "task3": 0.99
   },
-  "average": 0.9899999999999999
+  "average": 0.99
 }

inference.py

@@ -272,7 +272,7 @@ def main():
     print("="*60, file=sys.stderr)
     for k, v in scores.items():
         print(f"  {k}: {v:.4f}", file=sys.stderr)
-    avg = sum(scores.values()) / len(scores)
+    avg = round(sum(scores.values()) / len(scores), 4)
     print(f"  average: {avg:.4f}", file=sys.stderr)
     print("="*60, file=sys.stderr)