DataMaster Agent Automates Instruction Data Selection for LLMs

Current methods for selecting instruction data for large language model (LLM) training often require developers to manually inspect datasets and create specific rules for each application. This process is time-consuming and prone to errors because no single metric can universally apply to the complex, diverse nature of real-world data. To address this, researchers propose DataMaster, an Instruction Data Selection Agent. DataMaster shifts the paradigm from manual configuration to automated orchestration by interpreting a user's natural language description of their data needs. It then autonomously designs and applies the most effective data selection strategies. Extensive experiments across domains like mathematics, medicine, and code demonstrate DataMaster's superior performance. It consistently outperforms static baseline methods and, in many cases, even surpasses models trained on the entire data pool, simplifying data curation and enhancing LLM training efficiency.