arxiv preprint - Dataset Decomposition: Faster LLM Training with Variable Sequence Length Curriculum | AI Breakdown | Podwise