mediBuddy

Deepseek-R1 fine-tuned to be an efficient medical assistant chatbot

Fine-Tuning DeepSeek Model using LoRA

This repository provides a comprehensive guide to fine-tuning the DeepSeek model using LoRA (Low-Rank Adaptation) along with SFTTrainer. The process leverages parameter-efficient fine-tuning and mixed precision training to improve domain-specific performance while reducing memory usage and speeding up training.

---

Table of Contents

• Overview
• Prerequisites
• Fine-Tuning Steps
• Mixed Precision Training
• Differences in Model Output
• GPU Information
• Results and Monitoring

---

Overview

This repository details the process of creating an advanced medical assistant chatbot by fine-tuning the DeepSeek-R1 model. By leveraging LoRA for parameter-efficient fine-tuning and employing mixed precision training, MediBuddy is engineered to provide:

Key Features

• Domain-Specific Responses: Accurate, medically informed answers. Tailored outputs for medical queries.
• Enhanced Contextual Understanding: Maintains context across complex conversations, multi-turn conversations.
• Optimized Performance: Faster training and inference using FP16 mixed precision.
• Mixed Precision Training: Utilizes FP16 for improved computational efficiency.
• Parameter-Efficient Fine-Tuning: Implements LoRA to adjust only a subset of parameters.
• High Performance on Modern GPUs: Fine-tuned on an NVIDIA A100 for optimal throughput.

---

Medical Application and Disclaimer

MediBuddy has been specifically developed for medical applications, with an emphasis on medical imaging diagnostics and research. The model is fine-tuned on a curated dataset containing medical literature, imaging protocols, and domain-specific queries.

Disclaimer:
This chatbot is intended for research and academic purposes only. It should not be used as a substitute for professional medical advice, diagnosis, or treatment. Always consult a qualified healthcare professional for medical decisions.

---

The goal of this project is to fine-tune the DeepSeek model, adapting it for domain-specific tasks. By incorporating LoRA modifications and training on a custom dataset, the model produces more accurate and context-aware outputs compared to its pre-trained counterpart.

---

Prerequisites

• Python: 3.8+
• PyTorch: Latest version with GPU support
• Transformers: Latest release
• PEFT: For parameter-efficient fine-tuning
• Unsloth (fast_lora): For optimized LoRA kernels
• Weights & Biases (wandb): For experiment tracking
• Other dependencies: Listed in requirements.txt

---

Fine-Tuning Steps

• Load Base Model & Tokenizer
• Apply LoRA Configuration - train the target modules for generations query, key and value proj modules
• Configure Training Parameters with SFT trainer
• Start finetuning by running the trainer

---

Mixed Precision Training

Mixed precision training uses FP16 (half-precision floating point) arithmetic to reduce memory usage and accelerate computations. In this project, FP16 is enabled by setting fp16=True in the training configuration. Although BF16 is an alternative on some platforms, it has been disabled (bf16=False) to ensure compatibility with our custom unsloth kernels, thereby maintaining training stability.

---

Differences in Model Output

After fine-tuning, MediBuddy delivers significant improvements over the generic pre-trained model:

Aspect	Before Fine-Tuning	After Fine-Tuning (Deepseek-R1)
Response Specificity	Generic, broad responses	Tailored, medically accurate, and domain-specific responses
Contextual Understanding	Limited retention of context in complex queries	Enhanced context awareness for multi-turn medical dialogs
Technical Accuracy	Lower precision in handling specialized medical queries	Higher precision with accurate medical terminology and diagnostic details
Coherence	Basic coherence without specialized structure	Improved coherence with structured, relevant medical information

---

GPU Information

The fine-tuning process was executed on an NVIDIA A100 GPU (40GB), selected for its robust support for FP16 mixed precision training and high throughput, which substantially accelerates the training process.

---

Results and Monitoring

Training progress and performance metrics are tracked using Weights & Biases. For detailed metrics, visualizations, and performance comparisons, please refer to our Wandb Dashboard

Train results

GPU Power usage