diff --git a/dataflow/operators/core_text/generate/prompt_templated_qa_generator.py b/dataflow/operators/core_text/generate/prompt_templated_qa_generator.py
index 00b156c..bc639bf 100644
--- a/dataflow/operators/core_text/generate/prompt_templated_qa_generator.py
+++ b/dataflow/operators/core_text/generate/prompt_templated_qa_generator.py
@@ -107,9 +107,10 @@ def run(
                 key_dict[key] = val if pd.notna(val) else ""
                 
             prompt_text = self.prompt_template.build_prompt(need_fields, **key_dict)
-            
-            # 统一组装为基类所需的消息格式
-            conversations_list.append([{"role": "user", "content": prompt_text}])
+            if use_api_mode:
+                conversations_list.append([{"role": "user", "content": prompt_text}])
+            else:
+                conversations_list.append([{"from": "human", "value": prompt_text}])
 
         self.logger.info(
             f"Built {len(conversations_list)} prompts using fields: {need_fields}"
diff --git a/dataflow/operators/core_vision/generate/batch_vqa_generator.py b/dataflow/operators/core_vision/generate/batch_vqa_generator.py
index 398bb1c..0fe1de6 100644
--- a/dataflow/operators/core_vision/generate/batch_vqa_generator.py
+++ b/dataflow/operators/core_vision/generate/batch_vqa_generator.py
@@ -97,7 +97,11 @@ def run(self, storage: DataFlowStorage, input_prompts_key: str, input_image_key:
                 else:
                     # Local 模式（如 vLLM）通常需要手动在文本前拼接 <image> 占位符
                     img_tokens = "<image>" * len(image_path)
-                    conversation = [{"role": "user", "content": img_tokens + q}]
+                    # 根据 serving 模式选择对话格式
+                    if use_api_mode:
+                        conversation = [{"role": "user", "content": img_tokens + q}]
+                    else:
+                        conversation = [{"from": "human", "value": img_tokens + q}]
 
                 flat_conversations.append(conversation)
                 flat_images.append(image_path)
diff --git a/dataflow/operators/core_vision/generate/fix_prompted_vqa_generator.py b/dataflow/operators/core_vision/generate/fix_prompted_vqa_generator.py
index df4a22a..527e960 100644
--- a/dataflow/operators/core_vision/generate/fix_prompted_vqa_generator.py
+++ b/dataflow/operators/core_vision/generate/fix_prompted_vqa_generator.py
@@ -96,7 +96,10 @@ def run(self,
                 else:
                     content = ("<image>" * valid_media_count) + self.user_prompt
                     
-                conversations_list.append([{"role": "user", "content": content}])
+                if use_api_mode:
+                    conversations_list.append([{"role": "user", "content": content}])
+                else:
+                    conversations_list.append([{"from": "human", "value": content}])
                 
         elif has_videos:
             video_inputs_list = video_column
@@ -108,7 +111,10 @@ def run(self,
                 else:
                     content = ("<video>" * valid_media_count) + self.user_prompt
                     
-                conversations_list.append([{"role": "user", "content": content}])
+                if use_api_mode:
+                    conversations_list.append([{"role": "user", "content": content}])
+                else:
+                    conversations_list.append([{"from": "human", "value": content}])
 
         # 统一调用基类的消息生成接口
         outputs = self.serving.generate_from_input_messages(
diff --git a/dataflow/operators/core_vision/generate/prompt_templated_vqa_generator.py b/dataflow/operators/core_vision/generate/prompt_templated_vqa_generator.py
index 85d1d95..a85e123 100644
--- a/dataflow/operators/core_vision/generate/prompt_templated_vqa_generator.py
+++ b/dataflow/operators/core_vision/generate/prompt_templated_vqa_generator.py
@@ -128,7 +128,7 @@ def run(
         else:
             self.logger.info("Using local serving mode")
 
-        # 4. 构造多模态对话结构
+        # 4. 构造多模态对话结构（兼容 Local 和 API 模式）
         conversations_list = []
         image_inputs_list = None
         video_inputs_list = None
@@ -139,25 +139,29 @@ def run(
                 valid_media_count = len([p for p in paths if p])
                 
                 if use_api_mode:
+                    # API 模式：使用 OpenAI 格式
                     content = prompt_text
+                    conversations_list.append([{"role": "user", "content": content}])
                 else:
+                    # Local 模式：使用 ShareGPT 格式，并注入 <image> tokens
                     img_tokens = "<image>" * valid_media_count
                     content = f"{img_tokens}\n{prompt_text}" if img_tokens else prompt_text
+                    conversations_list.append([{"from": "human", "value": content}])
                     
-                conversations_list.append([{"role": "user", "content": content}])
-                
         elif has_videos:
             video_inputs_list = video_column
             for prompt_text, paths in zip(prompt_column, video_column):
                 valid_media_count = len([p for p in paths if p])
                 
                 if use_api_mode:
+                    # API 模式：使用 OpenAI 格式
                     content = prompt_text
+                    conversations_list.append([{"role": "user", "content": content}])
                 else:
+                    # Local 模式：使用 ShareGPT 格式，并注入 <video> tokens
                     vid_tokens = "<video>" * valid_media_count
                     content = f"{vid_tokens}\n{prompt_text}" if vid_tokens else prompt_text
-                    
-                conversations_list.append([{"role": "user", "content": content}])
+                    conversations_list.append([{"from": "human", "value": content}])
 
         # 5. 统一调用基类接口
         outputs = self.serving.generate_from_input_messages(
diff --git a/dataflow/operators/core_vision/generate/visual_reasoning_generator.py b/dataflow/operators/core_vision/generate/visual_reasoning_generator.py
index 5bc3232..3f62d1e 100644
--- a/dataflow/operators/core_vision/generate/visual_reasoning_generator.py
+++ b/dataflow/operators/core_vision/generate/visual_reasoning_generator.py
@@ -108,7 +108,10 @@ def run(
                 img_tokens = "<image>" * len(valid_img_paths)
                 content = f"{img_tokens}\n{q}" if img_tokens else q
 
-            flat_conversations.append([{"role": "user", "content": content}])
+            if use_api_mode:
+                flat_conversations.append([{"role": "user", "content": content}])
+            else:
+                flat_conversations.append([{"from": "human", "value": content}])
             flat_images.append(valid_img_paths)
             indices_to_generate.append(idx)
 
diff --git a/dataflow/operators/core_vision/generate/vlm_bbox_generator.py b/dataflow/operators/core_vision/generate/vlm_bbox_generator.py
index d70e929..76a3b6b 100644
--- a/dataflow/operators/core_vision/generate/vlm_bbox_generator.py
+++ b/dataflow/operators/core_vision/generate/vlm_bbox_generator.py
@@ -119,8 +119,11 @@ def run(self, storage: DataFlowStorage, input_image_key: str, input_kws_key: str
                 else:
                     img_tokens = "<image>" * len(img_path)
                     content = f"{img_tokens}\n{text_prompt}" if img_tokens else text_prompt
+                if use_api_mode:
+                    flat_conversations.append([{"role": "user", "content": content}])
+                else:
+                    flat_conversations.append([{"from": "human", "value": content}])
                 
-                flat_conversations.append([{"role": "user", "content": content}])
                 flat_images.append(img_path)
                 row_mappings.append({"row_idx": idx, "keyword": kw})
 
diff --git a/dataflow/operators/core_vision/refine/visual_dependency_refiner.py b/dataflow/operators/core_vision/refine/visual_dependency_refiner.py
index 46bb494..d1982e3 100644
--- a/dataflow/operators/core_vision/refine/visual_dependency_refiner.py
+++ b/dataflow/operators/core_vision/refine/visual_dependency_refiner.py
@@ -143,7 +143,10 @@ def run(self, storage: DataFlowStorage, input_list_key: str, input_image_key: st
                         img_tokens = "<image>" * len(image_path)
                         content_v = f"{img_tokens}\n{prompt_v}" if img_tokens else prompt_v
                         
-                    vis_conversations.append([{"role": "user", "content": content_v}])
+                    if use_api_mode:
+                        vis_conversations.append([{"role": "user", "content": content_v}])
+                    else:
+                        vis_conversations.append([{"from": "human", "value": content_v}])
                     vis_images.append(image_path)
                     vis_mappings.append({"row_idx": row_idx, "qa_idx": qa_idx, "expected": ans_v})
 
@@ -151,7 +154,10 @@ def run(self, storage: DataFlowStorage, input_list_key: str, input_image_key: st
                     q_t, ans_t = shuffle_options_logic(qa_item, add_none_option=False)
                     prompt_t = self.inst_template.format(q_t)
                     
-                    txt_conversations.append([{"role": "user", "content": prompt_t}])
+                    if use_api_mode:
+                        txt_conversations.append([{"role": "user", "content": prompt_t}])
+                    else:
+                        txt_conversations.append([{"from": "human", "value": prompt_t}])
                     txt_mappings.append({"row_idx": row_idx, "qa_idx": qa_idx, "expected": ans_t})
 
         # =========================================================
diff --git a/dataflow/operators/core_vision/refine/visual_grounding_refiner.py b/dataflow/operators/core_vision/refine/visual_grounding_refiner.py
index d97fcf5..b039f88 100644
--- a/dataflow/operators/core_vision/refine/visual_grounding_refiner.py
+++ b/dataflow/operators/core_vision/refine/visual_grounding_refiner.py
@@ -109,7 +109,10 @@ def run(self, storage: DataFlowStorage, input_list_key: str, input_image_key: st
                     img_tokens = "<image>" * len(image_path)
                     content = f"{img_tokens}\n{prompt_text}" if img_tokens else prompt_text
                 
-                flat_conversations.append([{"role": "user", "content": content}])
+                if use_api_mode:
+                    flat_conversations.append([{"role": "user", "content": content}])
+                else:
+                    flat_conversations.append([{"from": "human", "value": content}])
                 flat_images.append(image_path)
                 row_mappings.append({"row_idx": idx, "item": item})