DL_Project_2

Deep Learning Spring 2025: Finetuning with LoRA Parameter. Efficient Finetuning of LLMs with Low-Rank Adaption (LoRA) for News Data

LoRA Fine-Tuning with Hugging Face Transformers

This project explores Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning of large language models using Hugging Face's transformers and peft libraries. The notebook provides a complete pipeline for adapting a base model (such as GPT or LLaMA) using minimal additional parameters while preserving efficiency and performance.

📌 Objectives

• Understand and apply the LoRA technique for efficient fine-tuning
• Leverage Hugging Face's peft and transformers libraries
• Load a pre-trained language model and apply LoRA adaptation
• Train on a sample dataset
• Save and test the adapted model

---

📁 Project Structure

├── Final_LoRaFits_Project_2.ipynb # Main notebook with code and explanation

├── Final_submission.csv # csv file with output on test dataet submitted to kaggle

├── README.md # Project description (this file)

---

🔧 Requirements

Install the following Python libraries:

pip install transformers datasets peft accelerate

How to Run

1. Clone this repository or download the .ipynb file.

2. Install the dependencies listed above.

3. Open the notebook in Jupyter Notebook or JupyterLab:

jupyter notebook Final_LoRaFits_Project_2.ipynb

4. Follow each cell sequentially to:

• Load the base model

• Apply LoRA

• Prepare dataset

• Train the model

• Evaluate results

===== MODEL CONFIGURATION SUMMARY ===== Model architecture: RoBERTa with LoRA adapter

LoRA rank (r): 16

LoRA alpha: 32

Target modules: {'query', 'value', 'key'}

Total trainable parameters: 884,736

Validation accuracy: 0.9376

Test accuracy: 0.9404