feat: 增加测试结果;

main.py

@@ -208,7 +208,6 @@ def filter_rows_by_llm_result(
 			continue
 			continue
 
 
 		ground_truth_text = str(ground_truth).strip()
 		ground_truth_text = str(ground_truth).strip()
-		ground_truth_text = '合格'
 
 
 		try:
 		try:
 			for attempt in Retrying(wait=wait_fixed(1), stop=stop_after_attempt(max_retries), reraise=True):
 			for attempt in Retrying(wait=wait_fixed(1), stop=stop_after_attempt(max_retries), reraise=True):
@@ -231,7 +230,7 @@ def filter_rows_by_llm_result(
 						)
 						)
 						raise ValueError("llm result mismatch")
 						raise ValueError("llm result mismatch")
 					print(f"[llm_match] row={index}, attempt={attempt.retry_state.attempt_number},result={result_text}")
 					print(f"[llm_match] row={index}, attempt={attempt.retry_state.attempt_number},result={result_text}")
-					row['rsp'] = f'<think>{reasoning_content}</think>{result_text}'
+					row['rsp'] = f'<think>{reasoning_content}</think>{content}'
 		except Exception as exc:
 		except Exception as exc:
 			row['rsp'] = ''
 			row['rsp'] = ''
 			print(f"[llm_drop] row={index}, error={exc}")
 			print(f"[llm_drop] row={index}, error={exc}")