Code Contest

A repository for the course project of RL

Project Description

This project is a reinforcement learning based code contest system. The goal is to design a system that can automatically generate code for a given problem and evaluate the performance of the generated code. The system will be implemented with Python.

Project Structure

The project is organized into several folders:

• code: Contains code files generated for model evaluation (for future use).
• config: Contains configuration files for the model (for future use).
• data: Contains the data generated by process.py for training and testing the model.
• raw_data: Contains the raw data.
• results: Contains the results for platform submission.
• scripts: Contains utility functions used for data preprocessing, model training, and evaluation.
• tmp: Contains temporary cpp/exe files (cleaned up after evaluation).

Usage

To use the project, you need to install the required packages and run the following commands:

1. Install the required packages:

   pip install -r requirements.txt

2. Preprocess the raw data (only when data folder is empty or needs modification):

   Put your raw data in the raw_data folder.

   ---raw_data
    |--- 1_ShellGame
    | |--- data
    | | |--- 1.in
    | | |--- 1.out
    | | |--- ...
    | |--- 题面.md
    | 
    |--- 2_SleepCowHerding
    | |--- data
    | | |--- 1.in
    | | |--- 1.out
    | | |--- ...
    | |--- 题面.md
    | 
    ...
   

   Then run:

   python scripts/process_data.py

   This will generate the data in the data folder.

3. Generate code:

   python scripts/llm_coder.py --model_name "glm"

4. Run code:

   Utilize scripts/run_code.py to run the generated code.

5. Submit and evaluate code:

   python scripts/frontend.py <Competition ID>

   The competition ID can be found on the platform.

   Note: This API provide by the organizers is unbelivably slow.

Paper (Report)[https://github.com/Painkillerzzz/code_contest/blob/master/Reinforced_Learning_Project_Report.pdf]