training-scheduling

This paper analyzes curriculum learning in large language models by examining optimization dynamics across difficulty levels, proposing a new method called Transfer-aware Dynamic Curriculum Sampling (TDCS) that dynamically adjusts training data based on transfer relationships.