Update app.py

app.py

@@ -306,8 +306,40 @@ def two_stage_qa(question, candidate_paragraphs_str, max_seq_len_mc=512, max_seq
     if len(qa_features_dataset) == 0:
         return "錯誤: 無法為選定段落生成QA特徵 (可能段落太短或內容問題)。", f"選中的段落 (索引 {selected_idx}):\n{selected_paragraph}", "N/A"
 
-    # 創建 DataLoader
-    from transformers import default_data_collator # 需要導入
+    logger.info(f"--- In two_stage_qa, about to create DataLoader for question_id: {question_id} ---")
+    logger.info(f"Number of features in qa_features_dataset: {len(qa_features_dataset)}")
+
+    for i in range(len(qa_features_dataset)):
+        feature_item = qa_features_dataset[i]
+        logger.info(f"  Inspecting feature {i} from qa_features_dataset:")
+        for key_to_check in ["input_ids", "attention_mask", "token_type_ids"]:
+            if key_to_check not in feature_item:
+                logger.error(f"    !!!!!! CRITICAL: Key '{key_to_check}' is MISSING in feature {i}! Features: {feature_item.keys()}")
+                # 這是一個嚴重問題，會導致後續 collate 失敗
+                return f"錯誤: 特徵準備失敗，缺少 {key_to_check}", "N/A", "N/A"
+
+            val_list = feature_item[key_to_check]
+            if val_list is None:
+                logger.error(f"    !!!!!! CRITICAL: Key '{key_to_check}' in feature {i} is None!")
+                return f"錯誤: 特徵準備失敗，{key_to_check} 為 None", "N/A", "N/A"
+            
+            if not isinstance(val_list, list):
+                logger.error(f"    !!!!!! CRITICAL: Key '{key_to_check}' in feature {i} is not a list, but {type(val_list)}!")
+                return f"錯誤: 特徵準備失敗，{key_to_check} 不是列表", "N/A", "N/A"
+
+            if not val_list: # 如果列表為空
+                 logger.warning(f"    Feature {i} has an empty list for '{key_to_check}'. This might be okay if handled by collator for padding, but check if intended.")
+            
+            # 檢查列表內部元素
+            for elem_idx, elem in enumerate(val_list):
+                if elem is None:
+                    logger.error(f"    !!!!!! CRITICAL: Key '{key_to_check}' in feature {i} contains None at index {elem_idx}!")
+                    return f"錯誤: 特徵準備失敗，{key_to_check} 內部有 None", "N/A", "N/A"
+                if not isinstance(elem, int):
+                    logger.error(f"    !!!!!! CRITICAL: Key '{key_to_check}' in feature {i} contains non-integer {elem} (type: {type(elem)}) at index {elem_idx}!")
+                    return f"錯誤: 特徵準備失敗，{key_to_check} 內部有非整數", "N/A", "N/A"
+            logger.info(f"    Feature {i}, key '{key_to_check}' passed inspection. Length: {len(val_list)}")
+
     qa_dataloader = DataLoader(
         qa_features_dataset, collate_fn=default_data_collator, batch_size=8 # batch_size可以小一些
     )