https://github.com/nikisetti01/mtl-lora-for-pubmedqa-and-riddle

🚀 Fine-tuning LLaMA 1B for a medical chatbot using LoRA and a custom MTL-LoRA framework in PyTorch, enabling efficient multi-task learning for medical NLP! 🏥💡
https://github.com/nikisetti01/mtl-lora-for-pubmedqa-and-riddle

ai fine-tuning llm lora peft-fine-tuning-llm pytorch

Last synced: 2 months ago
JSON representation

🚀 Fine-tuning LLaMA 1B for a medical chatbot using LoRA and a custom MTL-LoRA framework in PyTorch, enabling efficient multi-task learning for medical NLP! 🏥💡

• Host: GitHub
• URL: https://github.com/nikisetti01/mtl-lora-for-pubmedqa-and-riddle
• Owner: nikisetti01
• Created: 2025-02-04T17:43:59.000Z (3 months ago)
• Default Branch: main
• Last Pushed: 2025-02-05T07:59:01.000Z (3 months ago)
• Last Synced: 2025-02-21T05:16:34.685Z (2 months ago)
• Topics: ai, fine-tuning, llm, lora, peft-fine-tuning-llm, pytorch
• Language: Jupyter Notebook
• Homepage:
• Size: 180 KB
• Stars: 0
• Watchers: 1
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Lora and MTL-Lora a new frontier of Multi-task  fine-tuning for LM

🚀 Medical Chatbot Fine-Tuning with LoRA & MTL-LoRA 🏥💡

🔍 Overview

This project pushes the boundaries of medical AI by fine-tuning the LLaMA 1B model using LoRA and our custom-built Multi-Task LoRA (MTL-LoRA) framework, designed from scratch in PyTorch.

📚 Phase 1: Fine-Tuning with PubMedQA

We enhance the model’s medical expertise with PubMedQA, leveraging:

⚡ LoRA: Efficient low-rank adaptation for faster, lightweight fine-tuning.

🛠️ Traditional Fine-Tuning: Updating only the last layer for controlled training.

📊 How We Evaluate

We compare three configurations: Base Model, LoRA-Tuned Model, and Traditionally Fine-Tuned Model using:

🎯 Perplexity

🏆 BLEU Score

📈 ROUGE Score

🤖 Phase 2: Custom Multi-Task LoRA (MTL-LoRA)

We built MTL-LoRA from scratch in PyTorch, allowing efficient multi-task learning across various medical NLP tasks in a single training pipeline. Inspired by cutting-edge research ([arXiv 2410.09437](https://arxiv.org/abs/2410.09437)), this approach ensures:

🚀 Seamless multi-task adaptation without retraining per task.

🔬 Enhanced generalization across diverse medical datasets.

💰 Reduced computational cost compared to full fine-tuning.

🌍 Join the Future of Medical AI!

Contribute, experiment, and push the boundaries of what’s possible in AI-driven healthcare! 🏥💙