LLMFineTuneDataFromCode

LLMFineTuneDataFromCode is a Python package designed to streamline the process of generating fine-tuning datasets for large language models (LLMs). It leverages code repositories and research papers to create datasets and provides tools to train and run inference on fine-tuned LLMs.

This package is tailored for labs and organizations aiming to extract insights from their existing resources for LLM training while ensuring flexibility and extensibility.

---

Features

1. Data Generation Pipeline

• Code-Based Dataset Generation:
  • The generateCodeFineTune.py script processes GitHub repositories to generate training datasets from code files.
  • It includes dependency analysis, Q&A generation, and code completion tasks.
• Research Paper Dataset Generation:
  • The createFinalDataOutput.py script processes research papers in PDF format to generate Q&A pairs.
  • Questions can be customized by editing the QUESTIONS list.

2. LLM Fine-Tuning Tools

• Training:
  • Use the train.ipynb script to fine-tune an LLM with the generated dataset.
  • Tested with Meta’s Llama 3.1 8B model (requires access from Huggingface).
• Inference:
  • Run inference on the fine-tuned model using the inference.ipynb script to interactively query your model.

---

Installation

1. Clone the repository:

   git clone https://github.com/Mogtaba-Alim/LLMFineTuneDataFromCode.git
   cd LLMFineTuneDataFromCode

2. Install the dependencies:

   pip install -r requirements.txt

3. Set up your environment variables:

   • Create a .env file and include your API keys (e.g., OpenAI or Anthropic) as follows:

     OPENAI_KEY=your_openai_api_key
     CLAUDE_API_KEY=your_claude_api_key
     HF_KEY=your_huggingface_key
     

4. Ensure GPU resources are available and compatible. The scripts are tested with NVIDIA A100 GPUs (40GB VRAM) and CUDA 12.6.

---

Usage

1. Generate a Fine-Tuning Dataset

From Code Repositories

python generateCodeFineTune.py

• Specify GitHub repository links in the script.

From Research Papers

python createFinalDataOutput.py

• Place your research papers (PDFs) in the ./Papers folder.
• Customize the questions in the QUESTIONS list as needed.

2. Fine-Tune an LLM

python train.ipynb

• Update the training script with your Huggingface model and dataset paths.

3. Run Inference

python inference.ipynb

• Interactively query your fine-tuned model.

---

Requirements

• Python 3.8+
• CUDA 12.6
• Huggingface Transformers
• PyTorch
• Other dependencies listed in requirements.txt

---

Notes

• GPU Resources: The package is optimized for high-memory GPUs like A100. Adjust parameters in train.ipynb for smaller GPUs.
• Access Requirements: You must have a Huggingface account and request access to the model being fine-tuned (e.g., Meta's Llama 3.1 8B).

---

Contribution

We welcome contributions to improve the package! Please submit a pull request or create an issue to share your ideas or report bugs.

---

License

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

---

Acknowledgements

This package was developed by Mogtaba Alim to facilitate efficient dataset generation and LLM fine-tuning. Special thanks to the open-source community for their invaluable tools and libraries.

---

Feel free to replace or expand on any section as needed for your specific requirements. Let me know if you'd like to refine this further!