core-set-selection

This paper proposes MADS, a method that leverages neural activation states from LLMs to select diverse core sets for instruction tuning, showing that a 15% subset can outperform full-dataset fine-tuning on multiple benchmarks.