LLaMA-Customer-Support-Assistant

Fine-tuned LLaMA 3.2 model for customer support using Unsloth optimization. Trained on Bitext's 27K customer service dataset. Optimized for M1 Mac.

Features

• 2x faster training with Unsloth optimization
• LoRA fine-tuning
• Memory-efficient for M1 Macs
• Automated customer support responses
• Order cancellation and status handling

Dataset

• Source: Bitext Customer Support Dataset
• CSV: 27K Customer Support Responses

Project Structure

LLaMA-Customer-Support-Assistant/
├── fine_tuning.ipynb          # Training notebook
├── test_model.ipynb           # Testing notebook
├── customer_support_model/    # Model outputs
├── dataset/                   # CSV dataset
└── README.md

Installation

# Create conda environment
conda create -n llm_env python=3.10
conda activate llm_env

# Install PyTorch for M1
conda install pytorch torchvision torchaudio -c pytorch

# Install dependencies
pip install transformers datasets accelerate peft
pip install unsloth

Usage

1. Start Jupyter:

jupyter notebook

2. Execute notebooks:

• Run fine_tuning.ipynb for training
• Run test_model.ipynb for testing

Model Details

• Base Model: unsloth/Llama-3.2-3B-Instruct
• Optimization: Unsloth + MPS backend
• Fine-tuning: LoRA
• Training Parameters:
  • Batch size: 1
  • Learning rate: 1e-4
  • Epochs: 1
  • Max sequence length: 256

Requirements

• M1/M2 Mac
• 8GB+ RAM
• Python 3.10+
• PyTorch with MPS support

Acknowledgments

• Unsloth for optimization
• Bitext for the dataset

License

MIT

Citation

@misc{bitext2023customer,
    title={Customer Support LLM Training Dataset},
    author={Bitext},
    year={2023},
    publisher={GitHub},
    url={https://github.com/bitext/customer-support-llm-chatbot-training-dataset}
}