MLE-agent

MLE-Agent is a paired LLM agent designed specifically for machine learning engineers and researchers, aiming to be an intelligent companion in AI engineering and research. Through autonomous execution and interactive collaboration, it helps users complete a variety of tasks from requirement understanding to model deployment.

Main Uses

MLE-Agent's core use is to assist machine learning engineering and AI research. It can automatically build ML/AI baseline solutions, participate in Kaggle competitions and complete tasks independently, and help improve existing projects through interactive chat.

Main Capabilities

• Autonomous Baseline Construction: Automatically build ML/AI baseline solutions based on user needs, which can be relatively vague, such as "I want to predict stock prices based on historical data."
• End-to-End ML Tasks: Independently participate in Kaggle competitions and complete the full process from data preparation to model training and debugging.
• arXiv and Papers with Code Integration: Access best practices and cutting-edge methods to provide references for code and research plans.
• Intelligent Debugging: Ensure code quality through automatic debugger-coder interaction.
• File System Integration: Efficiently organize project structure.
• Comprehensive Tool Integration: Includes AI/ML functions and MLOps tools to achieve seamless workflows.
• Interactive CLI Chat: Provides an easy-to-use chat interface to enhance project interaction.
• Intelligent Advisor: Provides personalized suggestions and recommendations for ML/AI projects.
• Weekly Reports: Automatically generate detailed weekly work reports, including development progress, communication records, reference materials, and to-do items.

How to Use

1. Installation: Clone the repository and create a virtual environment.

   git clone https://github.com/MLSysOps/MLE-agent.git
   cd MLE-agent
   uv venv .venv
   source .venv/bin/activate  # Linux/macOS
   pip install -e .
   

   After installation, a project directory will be created in the current path, and the project needs to be started in the project directory.

2. Interactive Chat: Start interactive chat in the terminal under the project directory.

3. Prototype ML Baseline: Use the MLE agent to prototype an ML baseline according to the given requirements, and test the model on the local machine.

4. Generate Work Report: Use the MLE agent to summarize the weekly report, then visit http://localhost:3000/ to generate the report locally.--start-date and --end-date are optional parameters. If omitted, a report for the last 7 days is generated by default. You need to replace <git email> with the Git email,<path_to_git_repo> is the local Git repository path.

5. Kaggle competitions: Use basic commands to start a Kaggle competition, or let the agent complete tasks without human interaction (datasets and submission files need to be prepared in advance). Before running, make sure you have joined the competition.

Applicable scenarios

• Machine learning engineers and researchers need to quickly build ML/AI baselines.
• Participate in Kaggle competitions and want to automate coding and debugging.
• Need access to arXiv and Papers with Code for research ideas.
• Want an intelligent advisor to provide project suggestions.
• Need to automatically generate weekly work reports.
• Want to improve existing ML projects through interactive chat.

Limitations

• The specific list of supported models includes OpenAI, Anthropic, Gemini, Ollama, etc., but a complete list of all models is not provided.
• Some features (such as cloud data integration, local RAG support, function zoo, etc.) are planned and may not yet be fully implemented.
• Using Kaggle mode requires the user to have already joined the competition and prepared datasets and submission files.
• Installation and use require some command-line operation experience.
• Details such as specific charges, pricing, and platform compatibility are not yet provided.