Update app.py

app.py

@@ -361,10 +361,56 @@ def two_stage_qa(question, candidate_paragraphs_str, max_seq_len_mc=512, max_seq
         logger.error(f"從 qa_features_dataset 選擇列時出錯: {e}. Features: {qa_features_dataset.features}")
         return f"錯誤: 準備模型輸入時出錯 (列選擇)。 Error: {e}", "N/A", "N/A"
 
-    logger.info(f"--- In two_stage_qa, about to create DataLoader for question_id: {log_question_id} ---")
-    
+    logger.info("--- 手動檢查 features_for_dataloader 以模擬 default_data_collator ---")
+    if len(features_for_dataloader) > 0:
+        # default_data_collator 會接收一個 features 列表，這裡我們模擬只有一個 feature 的情況
+        # 因為對於第一個 test_item，qa_features_dataset (以及 features_for_dataloader) 只有一行
+        
+        # features_list_for_collator 將是 [features_for_dataloader[0]]
+        # 如果 qa_batch_size > 1 且 features_for_dataloader 行數也 > 1，這裡會更複雜
+        # 但錯誤發生在第一個批次，所以檢查第一個特徵就夠了。
+        
+        single_feature_to_collate = features_for_dataloader[0] 
+        keys_to_tensorize_by_collator = ["input_ids", "attention_mask", "token_type_ids"]
+
+        for k_collate in keys_to_tensorize_by_collator:
+            if k_collate in single_feature_to_collate:
+                value_to_tensorize = single_feature_to_collate[k_collate]
+                logger.info(f"  準備轉換鍵 '{k_collate}' 的值: {str(value_to_tensorize)[:100]}...") # 打印部分值
+                if value_to_tensorize is None:
+                    logger.error(f"    !!!!!! CRITICAL (Pre-Collate): {k_collate} 的值是 None!")
+                    return f"錯誤: 預整理時發現 {k_collate} 為 None", "N/A", "N/A"
+                if not isinstance(value_to_tensorize, list):
+                    logger.error(f"    !!!!!! CRITICAL (Pre-Collate): {k_collate} 的值不是列表，類型為 {type(value_to_tensorize)}!")
+                    return f"錯誤: 預整理時發現 {k_collate} 不是列表", "N/A", "N/A"
+                if not value_to_tensorize: # 空列表
+                    logger.warning(f"    Pre-Collate: {k_collate} 的值是空列表。")
+                
+                problem_found_in_list = False
+                for elem_idx, elem_val in enumerate(value_to_tensorize):
+                    if elem_val is None:
+                        logger.error(f"    !!!!!! CRITICAL (Pre-Collate): {k_collate}[{elem_idx}] 是 None!")
+                        problem_found_in_list = True
+                        break
+                    if not isinstance(elem_val, int):
+                        logger.error(f"    !!!!!! CRITICAL (Pre-Collate): {k_collate}[{elem_idx}] 不是整數，值: {elem_val}, 類型: {type(elem_val)}!")
+                        problem_found_in_list = True
+                        break
+                if problem_found_in_list:
+                    return f"錯誤: 預整理時在 {k_collate} 內部發現問題", "N/A", "N/A"
+                
+                logger.info(f"    鍵 '{k_collate}' 的預整理檢查通過。")
+            else:
+                logger.warning(f"  鍵 '{k_collate}' 不在 features_for_dataloader[0] 中。")
+    else:
+        logger.error("features_for_dataloader 為空，無法進行手動檢查。")
+        return "錯誤: features_for_dataloader 為空", "N/A", "N/A"
+
+
     qa_dataloader = DataLoader(
-        qa_features_dataset, collate_fn=default_data_collator, batch_size=8 # batch_size可以小一些
+        features_for_dataloader, 
+        collate_fn=default_data_collator,
+        batch_size=8 # 或者 args.qa_batch_size
     )
 
     all_start_logits = []