Rethinking Data: Towards Better Performing Domain-Specific Small Language Models