Applied AI·Fine-tuning and customisation
you keep one small adapter per task sitting on one shared base model, and turn the rank up when the change you want will not fit.
Rank and adapters
Draft summary, pending review
Rank (r) sets the capacity of LoRA's matrices, commonly 8 to 64; higher captures more change and risks more overfitting. The trained result is an adapter you load alongside the base model, and you can keep one adapter per task on one shared base.