llama3.2-customer-support-qlora

Llama 3.2 Customer Support Assistant (QLoRA)

A QLoRA fine-tuned version of Meta Llama 3.2 3B Instruct designed to generate professional customer support responses for common e-commerce scenarios such as refunds, shipping issues, returns, order tracking, and product exchanges.

This model was developed as an educational and portfolio project demonstrating parameter-efficient fine-tuning (PEFT) using QLoRA.

---

Base Model

• Model: Meta Llama 3.2 3B Instruct
• Architecture: LlamaForCausalLM
• Fine-Tuning Method: QLoRA (PEFT)
• Quantization: 4-bit NF4 (BitsAndBytes)

---

Dataset

Fine-tuned using:

• bitext/Bitext-customer-support-llm-chatbot-training-dataset

The dataset contains customer support conversations covering common customer service scenarios.

---

Training Configuration

---

Training Results

Only 0.2848% of the model parameters were trained while keeping the remaining 99.7% frozen.

---

Adapter Information

This repository contains only the trained LoRA adapter weights.

To use this model, first load the base model (meta-llama/Llama-3.2-3B-Instruct) and then apply the adapter using the PEFT library.

---

Example

User

My package hasn't arrived.

Assistant

I'm sorry to hear that your package hasn't arrived. Could you please provide your order number or tracking number so I can check the shipping status for you?

---

Loading the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE_MODEL = "meta-llama/Llama-3.2-3B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL)

model = PeftModel.from_pretrained(
    base_model,
    "andref218/llama3.2-customer-support-qlora"
)

---

Intended Use

This model is intended for:

• Customer support chatbots
• Educational purposes
• Fine-tuning demonstrations
• PEFT / QLoRA experimentation

---

Limitations

• The model was fine-tuned on a customer support dataset and is not intended as a general-purpose assistant.
• Responses should be reviewed before deployment in production environments.
• Performance depends on the quality and scope of the fine-tuning dataset.

---

Technologies

• Transformers
• TRL
• PEFT
• QLoRA
• BitsAndBytes
• PyTorch
• Hugging Face Datasets
• Weights & Biases

---

Author

André Fonseca