I revisited my Pierre Joubert dataset and created another LoRA this time intentionally overbaking it.