QSGS Method for Efficient Instruction Tuning via Submodular Selection

The Query-based Submodular functions with Gradient-guided Synthesis (QSGS) method improves instruction tuning by selecting a diverse, non-redundant subset of training data. By treating selection as a submodular graph cut over gradient similarity, it maximizes performance under tight data budgets.