Tag
This paper introduces GLOBE, a trajectory-aligned coreset selection framework that uses gradient trajectories across multiple checkpoints and multi-order matching with structured sparse optimization to select compact, representative training subsets, outperforming existing methods on six benchmarks.
This paper proposes a submodular coreset selection method for LLM benchmarks that selects a subset of prompts without using model evaluation outcomes, achieving score preservation across 35 benchmarks and 18 LLMs.