Optimizing Small Language Models on a Free T4 GPU