Tag
The paper introduces MASS, a hierarchical data selection framework for LLM post-training that uses manifold and sparse feature coverage to select high-value data subsets, outperforming existing baselines in experiments.