🎬 Fine Tuned BERT with PEFT and LoRA

This project fine-tunes a BERT-based model for binary sentiment classification on the IMDB movie reviews dataset using Parameter-Efficient Fine-Tuning (PEFT) and Low-Rank Adaptation (LoRA) — enabling high performance with minimal computational overhead.

---

🧠 Introduction

The goal of this project is to adapt DistilBERT for sentiment analysis — classifying IMDB reviews as positive or negative.
Instead of fully fine-tuning the entire network, we use PEFT and LoRA to efficiently update only a small subset of parameters, achieving faster training and lower memory consumption.

---

🚀 Key Techniques

🔹 Parameter-Efficient Fine-Tuning (PEFT)

PEFT optimizes large language models by training only select parameters while keeping most of the model frozen.
This is particularly useful when:

• You have limited computational resources.
• You need to fine-tune large models quickly for new tasks.

🔹 Low-Rank Adaptation (LoRA)

LoRA is a specific PEFT method that injects trainable low-rank matrices into pre-trained model layers.
These low-rank updates are lightweight but powerful, allowing efficient adaptation without retraining the entire model.

LoRA Configuration:

Parameter	Value
r (rank)	4
lora_alpha	32
lora_dropout	0.01
target_modules	['q_lin']

Together, PEFT + LoRA enable faster convergence, lower memory footprint, and high accuracy even with small batch sizes.

---

⚙️ Model Training Pipeline

1️⃣ Data Preparation

• Load the IMDB dataset.
• Randomly sample training and validation subsets.
• Tokenize text using the BERT tokenizer.

2️⃣ Model Initialization

• Base model: distilbert-base-uncased
• Task: Sequence Classification with labels Positive and Negative.

3️⃣ Training Configuration

Hyperparameter	Value
Learning Rate	1e-3
Batch Size	4
Epochs	1
Evaluation Strategy	End of each epoch
Save Strategy	End of each epoch

4️⃣ Training

Model fine-tuned using the 🤗 Transformers Trainer API, integrated with PEFT for parameter-efficient updates.

---

🧩 Setup Instructions

1️⃣ Install Dependencies

pip install datasets evaluate transformers[sentencepiece]
pip install accelerate -U
pip install peft

2️⃣ Run the Script python

python FineTunedBERTwithPEFT&LoRA.py

3️⃣ Evaluate the Model

Use the validation set to compute metrics such as accuracy, precision, recall, and F1-score.

📊 Expected Outcome

• Improved accuracy on sentiment classification with reduced fine-tuned parameters
• Lower GPU memory consumption and faster training time compared to full fine-tuning
• Demonstrates practical integration of LoRA adapters into transformer-based architectures

🧾 References

• Hugging Face Transformers Documentation
• PEFT Library
• LoRA Paper – Low-Rank Adaptation of Large Language Models