Fine-Tuning a Small Model with LoRA and QLoRA

Fully fine-tuning a 7-billion-parameter model means updating all 7 billion weights and storing gradients and optimiser state for each: far beyond one consumer GPU. LoRA (Low-Rank Adaptation) freezes the original model and trains small add-on matrices instead, often well under 1% of the weights. QLoRA also loads the frozen model in 4-bit. Together they let you fine-tune on a single free or cheap GPU.

🎬 Teaching a giant a new trick