SupportLLM

Fine-tune a small instruction-tuned LLM for customer support response generation.

This project demonstrates:

• synthetic instruction dataset creation
• LoRA/QLoRA fine-tuning
• response quality and safety evaluation
• a Streamlit customer-support demo

Installation

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Prepare data

python scripts/prepare_dataset.py --input data/sample_support.jsonl --output data/processed/train.jsonl

Train

python scripts/train_qlora.py --config configs/qlora_config.yaml --train_file data/processed/train.jsonl

Inference

python scripts/inference.py --message "I was charged twice this month. Can you help?"

Demo

streamlit run app/streamlit_app.py