In this video, I walk you through a full, practical fine-tuning pipeline for DeepSeek R1 (the distilled LLaMA-8B variant) on custom-style instruction data—then I show you how to save that tuned model locally and actually use it. The key idea is simple: you don’t need to touch every parameter to get strong task adaptation. I use parameter-efficient fine-tuning (LoRA/PEFT) so you can train faster, cheaper, and with less GPU pain, while still getting meaningful behavior changes. I start by loading the DeepSeek R1 distilled model + tokenizer using Unsloth, because it’s built to make the fine-tuning loop efficient (better matrix ops, less overhead). Then I prep a GPT-4-generated instruction dataset (vicgalle/alpaca-gpt4) into a conversation-friendly format and run supervised fine-tuning using SFTTrainer. After training, the important system-design step is deployment: I package and save the fine-tuned model into Ollama so it lives on your machine and you can chat with it offline. By the end, you’ll have a local workflow: fine-tune → export → run in Ollama → test reasoning and responses end-to-end.