SakuraAI

An Earl McGowen Company

🎯 LoRA Fine-Tuning Demo

Technical demonstration of LoRA adapters for domain-specific chatbots
Training domain: Mental health research literature

🔬

Technical Demonstration: This showcases LoRA fine-tuning for domain-specific chatbots. Mental health research was chosen as example training data to demonstrate the technique. This is NOT a mental health service, medical tool, or substitute for professional healthcare.

Read the Blog Post

🎯 Technical Architecture

Demonstrating LoRA fine-tuning for domain specialization
Mental health = example subject matter only

Training Data (Example Domain)

425 mental health research papers from PubMed
Subject matter chosen to demonstrate domain specialization

Python PubMed API SQLite

⭐ LoRA Fine-Tuning (Key Technique)

Base Model: Mistral 7B (open-source LLM)

Method: LoRA adapters trained on mental health papers

Hardware: RTX 5070 Ti (16GB VRAM) with 4-bit quantization

💡 LoRA enables domain specialization without retraining the entire model—only ~0.1% of parameters are trained!

PyTorch LoRA Adapters 4-bit Quantization PEFT

Backend

Flask API with crisis detection and medical disclaimers

Flask Transformers CUDA

Frontend

SvelteKit with real-time chat and responsive design

SvelteKit TypeScript CSS Grid

Deployment

Linux desktop with ngrok tunnel for production access

Linux ngrok GPU Inference
📋 Technical Demo Notice: This is a machine learning demonstration. Mental health is only the example subject matter used for training data. This is NOT a mental health service or medical tool.

Try asking: