Large language models (LLMs), exemplified by ChatGPT, exhibit immense potential in specific software engineering tasks. Boasting a wealth of knowledge and versatile capabilities, these models are, however, plagued by high costs and significant response latency when applied to specialized domains. On the other hand, traditional small language models (SLMs), like BERT, are known for their efficiency. Yet, their capabilities are somewhat constrained, and they often fall short in comprehensively grasping professional knowledge and semantic nuances.
