Does anyone have experience with parameter efficient tuning methods while mitigating catastrophic forgetting for continual knowledge learning? Working on something, and could use a sounding board