Fine-Tuning Qwen1.5-0.5B with LoRA & 4-bit Quantization

This project fine-tunes the Qwen1.5-0.5B model using 4-bit quantization and Low-Rank Adaptation (LoRA) to efficiently adapt it for legal text generation.

Features

• Uses 4-bit quantization for reduced memory footprint.
• Implements LoRA for efficient fine-tuning.
• Loads and processes a legal dataset in JSON format.
• Splits dataset into training and validation sets.
• Fine-tunes with Trainer from Hugging Face.
• Saves and downloads the fine-tuned model.

Installation

Ensure the required libraries are installed:

pip install torch torchvision torchaudio
pip install --upgrade fsspec gcsfs
pip install transformers datasets accelerate peft bitsandbytes

Running the Fine-Tuning Script

1. Clone the repository:

git clone https://github.com/yourusername/qwen-fine-tune.git
cd qwen-fine-tune

2. Run the fine-tuning script:

python fine_tune_qwen.py

File Structure

/qwen-fine-tune
│── fine_tune_qwen.py   # Main script to fine-tune the model
│── legal_finetune_data.json # Legal dataset for training
│── qwen_fine_tuned/    # Directory for the fine-tuned model

Dataset Format

The dataset should be a JSON file structured as follows:

[
  {
    "prompt": "Provide details on legal section 1",
    "completion": "Section 1 covers the introduction to legal principles..."
  }
]

How It Works

1. Loads the Qwen1.5-0.5B model with 4-bit quantization.
2. Tokenizes the dataset for efficient processing.
3. Applies LoRA fine-tuning to optimize memory usage.
4. Uses Hugging Face’s Trainer to train the model.
5. Saves and downloads the fine-tuned model.

Training Configuration

Key training arguments:

• 3 epochs with fp16 for performance.
• Batch size: 1 with gradient accumulation (8 steps).
• Saves model every epoch.
• Logging every 10 steps.

Saving & Downloading the Model

Once training is complete, the model is saved and compressed:

zip -r qwen_fine_tuned.zip ./qwen_fine_tuned

You can download the model directly from Colab.

Contribution

Feel free to fork and contribute! Open issues for any questions or improvements.

License

MIT License

Acknowledgments

• Transformers for pre-trained models
• BitsAndBytes for quantization
• PEFT for efficient parameter fine-tuning
• This project utilizes datasets from another GitHub repository for model training. Special thanks to the original dataset creator. 🎖 [https://github.com/civictech-India/Indian-Law-Penal-Code-Json/tree/main]