Create eval.py

A Python file to confirm the answer

eval.py

@@ -0,0 +1,52 @@
+
+import pandas as pd
+import re
+import argparse
+
+# The above library is required
+
+def evaluate_numpuzzle(csv_path):
+    # Loading CSV (target: Ans, prediction: AI's Ans)
+    df = pd.read_csv(csv_path)
+    
+    total = len(df)
+    correct = 0
+    
+    print(f"--- NumPuzzle-Easy Evaluation Report ---")
+    print(f"Total Samples: {total}\n")
+
+    for index, row in df.iterrows():
+        target = str(row['target']).strip()
+        # Remove all except numbers (spaces, line breaks, characters) from the AI answer
+        prediction_raw = str(row['prediction'])
+        prediction_digits = re.sub(r'\D', '', prediction_raw)
+        
+        #[Judgment logic]
+        # 1. Because the answer is often written at the end of the sentence, get the "last 9 digits" of the extracted number sequence
+        # 2. However, if the total is less than 9 digits, it will be treated as it is (8 digits or less)
+        final_pred = prediction_digits[-9:] if len(prediction_digits) >= 9 else prediction_digits
+        
+        is_match = (final_pred == target)
+        
+        if is_match:
+            correct += 1
+        else:
+            # 失敗例のログ（デバッグ用：最初の10件のみ表示）
+            if correct < 5:
+                print(f"Sample {index} Failed:")
+                print(f"  Input Target: {target}")
+                print(f"  AI Extracted: {final_pred} (Raw: '{prediction_raw[:50]}...')")
+                print(f"  Reason: {'Leading zero lost or calculation error' if len(final_pred) < 9 else 'Value mismatch'}")
+                print("-" * 30)
+
+    accuracy = (correct / total) * 100
+    print(f"\nFinal Results:")
+    print(f"✅ Accuracy: {accuracy:.4f}%")
+    print(f"✅ Correct: {correct} / {total}")
+    print(f"----------------------------------------")
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--csv", type=str, required=True, help="Path to the prediction CSV file")
+    args = parser.parse_args()
+    evaluate_numpuzzle(args.csv)