Before this lesson, I didn’t really understand what LoRA meant.
It stands for: Low-Rank Adaptation
It means: the original model weight is frozen and smaller adaptations are added and so the model's behaviour can be adapted without touching the original. A bit like a sidecar on a motorbike.
The biggest thing I learned was this way, the LLM can be fine tuned for specific domains like legal, etc. And also adapting image models to a specific style (which is how I'd always thought about it)