If this type of fine-tuning severely harms the general-purpose capabilities of the model, why is it advantageous versus using a slimmer model more suited to your goal from the start? I guess I don't fully grok the paper.