update ds_assistant

modelscope_agent/agents/data_science_assistant.py

@@ -223,7 +223,7 @@
 
 Even if the code has been executed successfully, doesn't mean it's totally correct. You need to carefully \
 check the code logic to ensure the code can accomplish the task correctly. Ignore the warning messages. \
-You don't need to check the metrics of the model. you\
+You don't need to check the metrics of the model.
 
 these are the previous code blocks, which have been executed successfully in the previous jupyter notebook code blocks \
 {previous_code_blocks}
@@ -233,6 +233,9 @@
 - [your step by step thought], incorrect
 
 don't generate code , just give the reason why the code is correct or incorrect.
+
+## Attention
+don't use the word 'incorrect' in your step by step thought.
 """
 
 CHECK_DATA_PROMPT = """
@@ -315,12 +318,15 @@ def _update_plan(self, user_request: str, curr_plan: Plan = None) -> Plan:
         call_llm_success = False
         call_llm_count = 0
         tasks_text = ''
+        messages = [{
+            'role':
+            'user',
+            'content':
+            PLAN_TEMPLATE.format(
+                context='User Request: ' + user_request + '\n', )
+        }]
         while not call_llm_success and call_llm_count < 10:
-            resp = self._call_llm(
-                prompt=PLAN_TEMPLATE.format(
-                    context='User Request: ' + user_request + '\n', ),
-                messages=None,
-                stop=None)
+            resp = self._call_llm(prompt=None, messages=messages, stop=None)
             tasks_text = ''
             for r in resp:
                 tasks_text += r
@@ -568,7 +574,8 @@ def _judge_code(self, task, previous_code_blocks, code,
         if not call_llm_success:
             raise Exception('call llm failed')
         logger.info(f'judge result for task{task.task_id}: \n {judge_result}')
-        if 'incorrect' in judge_result:
+
+        if 'incorrect' in judge_result.split('\n')[-1]:
             success = False
             failed_reason = (
                 'Though the code executes successfully, The code logic is incorrect, here is the reason: '
@@ -604,12 +611,24 @@ def _run(self, user_request, save: bool = True, **kwargs):
                 while not success and code_counter < max_try:
                     code_execute_success = False
                     code_logic_success = False
+                    temp_code_interpreter = CodeInterpreter()
+
+                    temp_code_interpreter.call(
+                        params=json.dumps({
+                            'code':
+                            self._get_previous_code_blocks_without_outputs()
+                        }),
+                        nb_mode=True,
+                        silent_mode=True)
                     # generate code
                     code = self._generate_code(code_counter, task,
                                                user_request)
-
-                    code_execute_success, code_interpreter_resp = self.code_interpreter.call(
-                        params=json.dumps({'code': code}), nb_mode=True)
+                    code_execute_success, code_interpreter_resp = temp_code_interpreter.call(
+                        params=json.dumps({'code': code}),
+                        nb_mode=True,
+                        silent_mode=True)
+                    # 删除临时 jupyter环境
+                    temp_code_interpreter.terminate()
                     judge_resp = ''
                     if not code_execute_success:
                         logger.error(
@@ -631,10 +650,9 @@ def _run(self, user_request, save: bool = True, **kwargs):
                             result=code_interpreter_resp + '\n' + judge_resp,
                             is_success=False))
 
-                    if not success:
-                        # delete the last cell if the code execution failed
-                        del self.code_interpreter.nb.cells[-1]
-                    else:
+                    if success:
+                        self.code_interpreter.call(
+                            params=json.dumps({'code': code}), nb_mode=True)
                         task.code = code
                         task.result = code_interpreter_resp
                     code_counter += 1
@@ -650,7 +668,7 @@ def _run(self, user_request, save: bool = True, **kwargs):
                 else:
                     self.plan = self._update_plan(
                         user_request=user_request, curr_plan=self.plan)
-                    self.code_interpreter.nb.cells.clear()
+                    self.code_interpreter.reset()
             # save the plan into json file
             if save:
                 after_time = time.time()
@@ -662,10 +680,15 @@ def _run(self, user_request, save: bool = True, **kwargs):
                     'total_token': total_token,
                     'plan': self.plan.tasks
                 }
-                with open(
-                        dir_name + 'plan.json', 'w', encoding='utf-8') as file:
-                    file.write(
-                        json.dumps(plan_dict, indent=4, cls=TaskEncoder))
+                print(f'plan_dict: {str(plan_dict)}')
+                try:
+                    with open(
+                            dir_name + 'plan.json', 'w',
+                            encoding='utf-8') as file:
+                        file.write(
+                            json.dumps(plan_dict, indent=4, cls=TaskEncoder))
+                except Exception as e:
+                    print(f'json write error: {str(e)}')
 
         except Exception as e:
             logger.error(f'error: {e}')

modelscope_agent/tools/code_interpreter/code_interpreter_nb.py

@@ -45,6 +45,7 @@ def __init__(self, cfg={}):
         self.nb_client = NotebookClient(self.nb, timeout=180)
         self.console = Console()
         self.interaction = ''
+        self.silent_mode = False
         # timeout: int = 600
 
     def __del__(self):
@@ -138,8 +139,9 @@ def parse_outputs(self,
                 output_text = output['text']
             elif output['output_type'] == 'display_data':
                 if 'image/png' in output['data']:
-                    self.show_bytes_figure(output['data']['image/png'],
-                                           self.interaction)
+                    if not self.silent_mode:
+                        self.show_bytes_figure(output['data']['image/png'],
+                                               self.interaction)
 
             elif output['output_type'] == 'execute_result':
                 output_text = output['data']['text/plain']
@@ -227,6 +229,7 @@ def call(self,
              params: str,
              timeout: Optional[int] = 30,
              nb_mode: bool = False,
+             silent_mode: Optional[bool] = False,
              **kwargs) -> (bool, str):
         try:
             try:
@@ -249,7 +252,8 @@ def call(self,
                     )
             fixed_code = '\n'.join(fixed_code)
             if nb_mode:
-                result, success = self.run(code=fixed_code)
+                self.silent_mode = silent_mode
+                result, success = self.run(code=fixed_code, )
                 return success, result
         except Exception as e:
             return False, str(e)

modelscope_agent/tools/metagpt_tools/task_type.py

@@ -9,20 +9,20 @@
 - Remember to `import numpy as np` before using Numpy functions.
 - don't plot any columns
 - don't calculate correlations
+- Avoid showing all data, always use df.head() to show the first 5 rows.
 
 """
 
 # Prompt for taking on "data_preprocess" tasks
 DATA_PREPROCESS_PROMPT = """
 The current task is about data preprocessing, please note the following:
-- Monitor data types per column, applying appropriate methods.
 - make sure train and test data MUST have the same columns except for the label column.
+- Remove ID columns if exist.
 - Handle missing values with suitable strategies.
-- Ensure operations are on existing dataset columns.
-- Avoid writing processed data to files.
 - Avoid any change to label column, such as standardization, etc.
 - Each step do data preprocessing to train, must do same for test separately at the same time.
 - Always copy the DataFrame before processing it and use the copy to process.
+- Avoid writing processed data to files.
 """
 
 # Prompt for taking on "feature_engineering" tasks
@@ -31,9 +31,9 @@
 - Avoid creating redundant or excessively numerous features in one step.
 - Each feature engineering operation performed on the train set must also applies to the \
 test separately at the same time.
-- Avoid using the label column to create features, except for cat encoding.
 - Use the data from previous task result if exist, do not mock or reload data yourself.
 - Always copy the DataFrame before processing it and use the copy to process.
+- Use Label Encoding for non-numeric columns, avoid One-Hot Encoding.
 """
 
 # Prompt for taking on "model_train" tasks
@@ -42,6 +42,7 @@
 - If non-numeric columns exist, perform label encode together with all steps.
 - If the model caused timeout error, please don't use this model again.
 - Use the data from previous task result directly, do not mock or reload data yourself.
+- Never save the model to a file.
 """
 
 # Prompt for taking on "model_evaluate" tasks

modelscope_agent/tools/metagpt_tools/tool_recommend.py

@@ -18,9 +18,20 @@
 - You can utilize pre-defined tools in any code lines from 'Available Tools' in the form of Python class or function.
 - You can freely combine the use of any other public packages, like sklearn, numpy, pandas, etc..
 
+
 ## Available Tools:
 Each tool is described in JSON format. When you call a tool, import the tool from its path first.
 {tool_schemas}
+
+## Attention
+Ensure that the tool is imported from the correct path, the tool path is \
+./modelscope_agent/tools/metagpt_tools/libs/xxx.py
+if you want to use the tool in your code， you need to import the tool first, like this:
+```python
+from modelscope_agent.tools.metagpt_tools.libs.xxx import ToolName
+```
+
+
 """
 
 TOOL_RECOMMENDATION_PROMPT = """