From Text to SQL: Fine-Tuning Phi-4-Mini-Instruct with LoRA and PyTorch (Deep Dive)

In this deep-dive video, I’ll walk through the full workflow of LoRA fine-tuning the Phi-4-Mini-Instruct model to improve natural-language-to-SQL accuracy. You’ll learn how to: • What fine-tuning is, how it compares to RAG (Retrieval-Augmented Generation), and when to use each • Coding a PyTorch fine-tuning script for small models • Formatting custom training data for LLMs • Running fine-tuning jobs in a Docker container • Merging fine-tuned adapters into a single model •Quantizing the model to run efficiently on edge or local devices • Testing your fine-tuned model on a desktop with LM Studio The demonstration runs on an NVIDIA DGX Spark, but the methods apply to any CUDA-based environment. If you’re a developer or AI engineer who wants to get hands-on with customizing small language models, this deep-dive is for you. Link to Github repo with scripts and data used in the video: https://github.com/robkerr/robkerrai-demo-code/tree/main/dgx-spark-lora-finetune-phi3 Link to NVIDIA Spark DGX Training with PyTorch Playbook: https://build.nvidia.com/spark/pytorch-fine-tune 00:00 - Introduction 03:07 - Training Data 05:23 - Training Script 11:18 - Docker Startup 15:37 - Python Env Setup – Running LoRA Fine-Tuning 16:47 - Training Script 21:11 - Merging Adapters 23:03 - Llama.cpp Setup 24:36 - Creating a gguf 25:33 - Quantizing the Model 29:05 - Testing with Llama.cpp 33:00 - Testing in LM Studio 36:55 - Summary