DataMaster: AI Agent for Automated Instruction Data Selection

A recent study published on arXiv (2608.10579) presents DataMaster, an Instruction Data Selection Agent designed to automate the selection of instruction data for AI models. The authors contend that current methods for selecting instruction data depend on various metrics; however, the intricate nature of real-world datasets means that no single metric can effectively generalize across all situations. Typically, developers must manually review data and create heuristic rules for each application, a process that is both labor-intensive and prone to errors. DataMaster transitions from manual setup to automated management by understanding user intent through natural language and independently generating optimal selection strategies. This innovation streamlines data curation and alleviates the need for manual strategy development. Comprehensive testing in math, medical, and coding fields indicates that DataMaster frequently outperforms static baselines and exceeds full-pool training in many instances. The research is authored by a team of researchers and can be accessed on arXiv.