Merge branch 'optimize-mle-bench-prompt' into 'sela'

Optimize mle bench prompt

See merge request agents/exp_optimizer!29

expo/data/custom_task.py

@@ -21,11 +21,11 @@ COMPETITION INSTRUCTIONS
 {task_description}
 
 ## More Instructions
-- output_dir: {output_dir}
-- Besides `submission.csv`, you should also save your output in the output directory.
-- You should split the training data into train and dev set.
+- You should split the training data into train and dev set with a seed of 42.
 - You should use the dev set to improve your model. Print the final dev set score after training.
-- Save the prediction results of BOTH the dev set and test set in `dev_predictions.csv` and `test_predictions.csv` respectively in the output directory. They should be in the same format as the `submission.csv`.
+- output_dir: {output_dir}
+- Besides `submission.csv`, you should also save your `test_predictions.csv` and `dev_predictions.csv` in the output directory.
+- Note that `test_predictions.csv` should be identical to `submission.csv`.
 - Perform data analysis, data preprocessing, feature engineering, and modeling to predict the target. {special_instruction}
 **Do not make any plots or visualizations.**
 """

expo/insights/instruction_generator.py

@@ -8,7 +8,7 @@ from expo.utils import clean_json_from_rsp, load_data_config, mcts_logger
 from metagpt.llm import LLM
 from metagpt.schema import Message
 
-REFLECTION_SYSTEM_MSG = "As a Kaggle grandmaster participating in a competition, you need to analyze your experience and propose evolutionary points that are more likely to improve the performance of baseline code."
+REFLECTION_SYSTEM_MSG = "As a Kaggle Grandmaster competing in a challenge, your task is to suggest potential evolutionary improvements that could enhance the performance of the baseline code."
 
 CHANGE_INSTRUCTION = """
 # Original instruction
@@ -17,7 +17,9 @@ CHANGE_INSTRUCTION = """
 # Insights
 {insights}
 
-Rewrite the original instruction according to the insights
+Rewrite the original instruction according to the insights 
+(If the original instruction involves splitting the data, ensure that your insights are integrated with the data split instructions, 
+rather than replacing them.)
 
 # Expected Output Hard Format
 ```json

expo/scripts/visualize_experiment.py

@@ -20,4 +20,4 @@ if __name__ == "__main__":
     root = mcts.root_node
     G = nx.DiGraph()
     build_tree_recursive(G, "0", root)
-    visualize_tree(G, save_path="results/tree.png")
+    visualize_tree(G, save_path=f"results/{args.task}-tree.png")