CoLLM

An Automated, End-to-End Deep Learning Toolbox for Collider Physics Analysis

CoLLM (Collider LLM) is an intelligent code generation tool that automates the creation of executable Python analysis scripts for LHCO (Les Houches Collider Olympics) files produced by fast detector simulations like Delphes. Simply describe your physics analysis in natural language, and CoLLM generates validated, runnable code.

✨ Key Features

• 🤖 LLM-Powered Code Generation — Leverages state-of-the-art code models (Qwen, DeepSeek) to generate physics analysis code
• 🔄 Automatic Error Correction — Self-healing code with automatic bug detection and fixing
• 🖥️ Dual Interface — Choose between Terminal UI (TUI) or Streamlit-based Graphical UI (GUI)
• ⚡ GPU Acceleration — Full support for CUDA (NVIDIA) and MPS (Apple Silicon)
• 📊 Built-in Validation — Syntax checking and pattern validation before execution
• 🔌 API Support — Use local models or HuggingFace Inference API

📦 Installation

pip install collm-hep

This installs everything you need, including the GUI.

🚀 Quick Start

Python API

from collm import generate_lhco_code

# Using local model
code = generate_lhco_code(
    user_input_path="user_input.txt",
    output_path="generated_analysis.py",
    model_id="Qwen/Qwen2.5-Coder-14B-Instruct"
)

# Using HuggingFace API
code = generate_lhco_code(
    user_input_path="user_input.txt",
    output_path="generated_analysis.py",
    model_id="Qwen/Qwen2.5-Coder-14B-Instruct",
    use_api=True,
    api_key="your_hf_api_key"
)

Command Line

# Generate code from input file
collm --input user_input.txt --output analysis.py

# Use HuggingFace API
collm --input user_input.txt --output analysis.py --api --api-key YOUR_KEY

# Run with configuration file
collm --config config.yml

# Launch GUI
collm --gui

User Input Format

Create a user input file with three sections:

[SELECTION_CUTS]
- Select electrons with pT > 10 GeV and |eta| < 2.5
- Select muons with pT > 10 GeV and |eta| < 2.4
- Require at least two leptons
- Require at least two jets

[PLOTS_FOR_VALIDATION]
- Plot the missing energy distribution
- Plot the invariant mass of leading and subleading leptons
- Normalize all histograms to one

[OUTPUT_STRUCTURE]
- Save the produced histograms into png with dpi=150
- Print summary statistics
- Print the number of events before and after selection cuts

⚙️ Configuration

Create a YAML configuration file for batch processing:

Output_dir: "./output/"
DEFAULT_MODEL: "Qwen/Qwen2.5-Coder-14B-Instruct"
MAX_RETRIES: 3
Input_file: "./data/signal.lhco"
User_input: "./templates/user_input.txt"
Use_api: False
Api_key: "your_huggingface_api_key"

🤖 Supported Models

📚 LHCO File Format Reference

Particle Type Codes:

• 0 = Photon
• 1 = Electron
• 2 = Muon
• 3 = Tau
• 4 = Jet
• 6 = MET

🔧 API Reference

Main Functions

from collm import generate_lhco_code, fix_code

# Generate LHCO analysis code
code = generate_lhco_code(
    output_path: str,
    model_id: str = "Qwen/Qwen2.5-Coder-14B-Instruct",
    user_input_path: Optional[str] = None,
    user_input_text: Optional[str] = None,
    system_prompt_path: Optional[str] = None,
    use_api: bool = False,
    api_key: Optional[str] = None
) -> Optional[str]

# Fix buggy code
fixed_code = fix_code(
    code: str,
    error: str,
    model_id: str = "Qwen/Qwen2.5-Coder-14B-Instruct",
    use_api: bool = False,
    api_key: Optional[str] = None
) -> str

Configuration Class

from collm.utils.config import CoLLMConfig

# Load from YAML
config = CoLLMConfig.from_yaml("config.yml")

# Access configuration
print(config.model_id)
print(config.output_dir)

# Save to YAML
config.to_yaml("new_config.yml")

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📖 Citation

If you use CoLLM in your research, please cite:

@software{collm2025,
  title = {CoLLM: An Automated Deep Learning Toolbox for Collider Physics Analysis},
  year = {2025},
  url = {https://github.com/yourusername/CoLLM}
}