subtopic

LoRA and QLoRA in simple words

Learn why full fine-tuning is costly (billions of weights) and how LoRA solves it: freeze the model and train small add-on matrices (often under 1% of the weights). QLoRA also loads the frozen model in 4-bit so a 1–3B model trains on a free 16 GB GPU. Key settings: rank r (8–16), alpha (16–32), which layers get adapters.