Merge branch 'expo' into 'sela'

open-source sela

See merge request agents/exp_optimizer!28

metagpt/actions/di/write_analysis_code.py

@@ -6,8 +6,6 @@
 """
 from __future__ import annotations
 
-import json
-
 from metagpt.actions import Action
 from metagpt.prompts.di.write_analysis_code import (
     CHECK_DATA_PROMPT,
@@ -30,9 +28,10 @@ class WriteAnalysisCode(Action):
         )
 
         rsp = await self._aask(reflection_prompt, system_msgs=[REFLECTION_SYSTEM_MSG])
-        reflection = json.loads(CodeParser.parse_code(block=None, text=rsp))
-
-        return reflection["improved_impl"]
+        # reflection = json.loads(CodeParser.parse_code(block=None, text=rsp))
+        # return reflection["improved_impl"]
+        reflection = CodeParser.parse_code(block=None, text=rsp)
+        return reflection
 
     async def run(
         self,

metagpt/prompts/di/write_analysis_code.py

@@ -40,15 +40,17 @@ Tests failed:
 assert add(1, 2) == 3 # output: -1
 assert add(1, 3) == 4 # output: -2
 
-[reflection on previous impl]:
+[reflection on previous impl]
 The implementation failed the test cases where the input integers are 1 and 2. The issue arises because the code does not add the two integers together, but instead subtracts the second integer from the first. To fix this issue, we should change the operator from `-` to `+` in the return statement. This will ensure that the function returns the correct output for the given input.
 
-[improved impl]:
+[improved impl]
+```python
 def add(a: int, b: int) -> int:
    """
    Given integers a and b, return the total value of a and b.
    """
    return a + b
+```
 '''
 
 REFLECTION_PROMPT = """
@@ -60,17 +62,17 @@ Here is an example of debugging with reflection.
 [context]
 {context}
 
-[previous impl]:
+[previous impl]
 {previous_impl}
 
 [instruction]
 Analyze your previous code and error in [context] step by step, provide me with improved method and code. Remember to follow [context] requirement. Don't forget to write code for steps behind the error step.
-Output a json following the format:
-```json
-{{
-    "reflection": str = "Reflection on previous implementation",
-    "improved_impl": str = "Refined code after reflection.",
-}}
+Output in the following format:
+[reflection on previous impl]
+...
+[improved impl]:
+```python
+# your code
 ```
 """
 

metagpt/prompts/task_type.py

@@ -11,7 +11,7 @@ The current task is about data preprocessing, please note the following:
 - Monitor data types per column, applying appropriate methods.
 - Ensure operations are on existing dataset columns.
 - Avoid writing processed data to files.
-- Avoid any change to label column, such as standardization, etc.
+- **ATTENTION** Do NOT make any changes to the label column, such as standardization, etc.
 - Prefer alternatives to one-hot encoding for categorical data.
 - Only encode or scale necessary columns to allow for potential feature-specific engineering tasks (like time_extract, binning, extraction, etc.) later.
 - Each step do data preprocessing to train, must do same for test separately at the same time.
@@ -25,8 +25,8 @@ The current task is about feature engineering. when performing it, please adhere
 - Use available feature engineering tools if they are potential impactful.
 - Avoid creating redundant or excessively numerous features in one step.
 - Exclude ID columns from feature generation and remove them.
-- 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.
+- Each feature engineering operation performed on the train set must also applies to the dev/test separately at the same time.
+- **ATTENTION** Do NOT use 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.
 """
@@ -34,6 +34,10 @@ The current task is about feature engineering. when performing it, please adhere
 # Prompt for taking on "model_train" tasks
 MODEL_TRAIN_PROMPT = """
 The current task is about training a model, please ensure high performance:
+- For tabular datasets - you have access to XGBoost, CatBoost, random forest, extremely randomized trees, k-nearest neighbors, linear regression, etc.
+- For image datasets - you have access to Swin Transformer, ViT, ResNet, EfficientNet, etc.
+- For text datasets - you have access to Electra, DeBERTa, GPT-2, BERT, etc.
+- Avoid the use of SVM because of its high training time.
 - Keep in mind that your user prioritizes results and is highly focused on model performance. So, when needed, feel free to use models of any complexity to improve effectiveness, such as XGBoost, CatBoost, etc.
 - If non-numeric columns exist, perform label encode together with all steps.
 - Use the data from previous task result directly, do not mock or reload data yourself.

metagpt/roles/role.py

@@ -478,10 +478,10 @@ class Role(SerializationMixin, ContextMixin, BaseModel):
 
     async def _plan_and_act(self) -> Message:
         """first plan, then execute an action sequence, i.e. _think (of a plan) -> _act -> _act -> ... Use llm to come up with the plan dynamically."""
-
-        # create initial plan and update it until confirmation
-        goal = self.rc.memory.get()[-1].content  # retreive latest user requirement
-        await self.planner.update_plan(goal=goal)
+        if not self.planner.plan.goal:
+            # create initial plan and update it until confirmation
+            goal = self.rc.memory.get()[-1].content  # retreive latest user requirement
+            await self.planner.update_plan(goal=goal)
 
         # take on tasks until all finished
         while self.planner.current_task: