Fetching the paper…

Dataset Decomposition: Faster LLM Training with Variable Sequence Length Curriculum · Around