Download and open the ipynb file locally, or else open the PDF

Instruction Fine-Tuning Phi-2 using QLoRA for Amazon Product Description,Name Generation

This repository demonstrates how to fine-tune the Phi-2 language model on custom Amazon product data using QLoRA (Quantized Low-Rank Adaptation). The goal is to efficiently adapt a large language model to generate product names and descriptions while minimizing GPU memory usage.

---

Project Overview

Product title and description generation is a practical text generation task in e-commerce. In this project, the Phi-2 model is instruction fine-tuned on structured product data using QLoRA, enabling efficient training by combining low-bit quantization with LoRA adapters. The notebook focuses on building a scalable and memory-efficient fine-tuning pipeline rather than full model retraining.

---

Dataset

• Custom dataset containing:
  • Amazon product names
  • Corresponding product descriptions
• Data is formatted into instruction-style or prompt–completion pairs suitable for causal language modeling
• Used for supervised fine-tuning of Phi-2

---

Model Used

Phi-2

• Lightweight large language model developed by Microsoft
• Decoder-only transformer architecture
• Suitable for instruction tuning and text generation tasks

---

Fine-Tuning Approach

QLoRA (Quantized Low-Rank Adaptation)

• Base model weights loaded in low-bit precision using bitsandbytes
• Trainable LoRA adapters injected into selected transformer layers (Wqkv,fc1,fc2)
• Enables fine-tuning large models on limited GPU memory
• Original model weights remain frozen

---

Workflow

1. Environment Setup

   • PyTorch
   • Hugging Face Transformers, Datasets, and PEFT
   • BitsAndBytes for low-bit quantization
   • Accelerate for efficient training

2. Data Preparation

   • Loading custom Amazon product dataset
   • Formatting data into prompt–target text pairs
   • Tokenization for causal language modeling

3. Model Loading

   • Phi-2 loaded with quantization configuration
   • QLoRA adapters configured using PEFT

4. Fine-Tuning

   • Supervised fine-tuning using Hugging Face Trainer
   • Training only LoRA parameters
   • Efficient memory usage suitable for Colab GPUs

5. Evaluation

   • Monitoring training and validation loss
   • Qualitative inspection of generated product text

6. Model Saving

   • Saving LoRA adapters
   • Option to merge adapters with base model for inference

---

Results

• The fine-tuned model learns domain-specific patterns in Amazon product data
• Generates coherent product names and descriptions aligned with training format
• Demonstrates effective adaptation using QLoRA with minimal GPU memory overhead

---

Technologies Used

• Python
• PyTorch
• Hugging Face Transformers
• Hugging Face PEFT
• BitsAndBytes
• Accelerate
• Google Colab (GPU)

---