LLMs are limited by the amount of memory you have, so you end up having to pick and choose what the model is trained on. LoRA is the popular customizing approach which can take an existing train model and then apply biases to it for a specific context. This avoids having to train a model from scratch for a specific context.