MM-LIMA: Less Is More for Alignment in Multi-Modal Datasets