APICoder: A Multi-Role Large Language Model Framework for API Service Call Code Generation

With the rapid advancement of large language models (LLMs) in code generation, the task of API service invocation code generation faces increasing challenges such as inaccurate context understanding, improper API selection, and outdated data. To address these issues, this paper constructs a high-quality dataset, APIDataset, and fine-tunes the DeepSeek-Coder-V2-Lite-Instruct model based on this dataset. We further integrate a retrieval-augmented generation (RAG) approach to enhance the model's contextual comprehension capabilities. Moreover, we propose APICoder, a multi-role collaborative framework that optimizes different stages of code generation, significantly improving the quality of generated API invocation code. Finally, we develop the APITest benchmark and conduct two rounds of evaluation combining both API-specific and general code generation test sets. Experimental results demonstrate that APICoder achieves superior performance in both API invocation and general code generation tasks, validating its effectiveness and practical potential.