小模型也能当好老师，强教师不总有效。
Strong Teacher Not Needed? On Distillation in LLM Pretraining
- 用混合损失让小模型指导大模型，效果优于传统强教师。
- 教师过强反而降低效果，参数或训练量过多会饱和甚至变差。
- 提升泛化能力更明显，适合做下游任务的预训练。
知识蒸馏通常假设强教师能教出好学生。本文通过调整模型架构大小和训练数据量，构建了强-弱、同级、弱-强三类师生关系，研究其在大语言模型预训练中的有效性。结果发现，教师无需强大：通过合理混合语言建模与蒸馏损失，即使小而欠训的教师也能提升大模型性能。同时，教师过强并不总有益：增加参数或训练样本可能使蒸馏收益饱和甚至逆转。此外，蒸馏更显著提升模型在分布外和下游任务的表现，而非仅优化领域内拟合。这些发现挑战了蒸馏需强教师的普遍认知。
原文摘要 · Abstract (English)
Knowledge distillation generally assumes a strong-to-weak relationship where stronger teachers yield better students. In this work, we examine this assumption about distillation in large language model pretraining. By varying architecture sizes and training token budgets, we create strong-to-weak, same-level, and weak-to-strong teacher-student relationships, and study distillation's effectiveness under each. We find that the teacher need not be strong: with proper mixing of the language modeling and knowledge distillation losses, even small and undertrained teachers improve larger students. At the same time, a stronger teacher is not always better: pushing the teacher further, through more parameters or more training tokens, can saturate or even reverse the distillation gains. We further observe that distillation improves generalization (out-of-distribution and downstream performance) more readily than in-domain fitting. Together, these results challenge the common belief that distillation pretraining always requires a strong teacher.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。