LoRA Fine Tuning Tutorial: Fine-Tune IBM Granite 4.2 in 20 Minutes

A LoRA fine tuning tutorial on a real run: six and a half minutes of training on one rented NVIDIA A100 taught IBM Granite 4.2 to answer "Hello" with "Ahoy there, matey." It explains fine-tuning and LoRA (low-rank adaptation), which freezes the model's 8 billion weights and trains a small add-on about 0.6% of its size, then walks through the training loop, the 22 GB memory peak and the cost at RunPod's rates. A before-and-after test of five everyday questions goes from zero pirate words to 103, and the video ends on when a plain prompt is enough and when fine-tuning is worth it. Chapters 0:00 Fine-Tuning IBM Granite Model 0:37 Choosing a Small Open-Weight Model 0:58 Renting a GPU and Training Data 1:25 LoRA Versus Full Retraining 2:25 Training Loop and Loss Explained 2:53 Running the Job in Terminal 4:46 Testing Before and After Training 6:11 Prompting Versus Fine-Tuning Guide Details from this video LoRA targets: q, k, v, o and the three MLP matrices in all 40 layers, rank 16 Merge the adapter into the model (PEFT, no added inference cost): merge_and_unload() — https://huggingface.co/docs/peft/main... Prompt-only baseline: system prompt "answer like a pirate" Sources Hugging Face, IBM Granite 4.2 8B model card: https://huggingface.co/ibm-granite/gr... Hugging Face, pirate-ultrachat-10k dataset: https://huggingface.co/datasets/wingl... arXiv, LoRA: Low-Rank Adaptation of Large Language Models: https://arxiv.org/abs/2106.09685 RunPod, GPU pricing: https://www.runpod.io/gpu-instance/pr... OpenAI supervised fine-tuning guide: https://developers.openai.com/api/doc... LoRA Without Regret (Thinking Machines Lab): https://thinkingmachines.ai/blog/lora/ OpenAI model optimization guide: https://developers.openai.com/api/doc... Footage NVIDIA A100 80GB, from HPCWH's "NVIDIA A100 80GB Tensor Core GPU Unboxing #nvidia #graphicscard #ai #tensorflow #datacentre":    • NVIDIA A100 80GB Tensor Core GPU Unboxing ...   #LoRA #IBMGranite #AIExplained — The Token Desk · AI news and explainers Subscribe:    / @thetokendesk