Patching open-vocabulary models by interpolating weights