Knowledge distillation improves small language models efficiently
KDFP: A first-principles approach to knowledge distillation in large language models
Machine Learning
Summary
Smaller language models often need help learning from bigger ones to work better on devices with limited power and privacy needs. The authors studied how to teach these smaller models by copying knowledge from bigger models more effectively. They created a new method called KDFP that learns general knowledge better than past approaches while also using less training resources. This method could help make smart language technology more accessible for things like phones or other devices that can’t handle large models. The authors tested KDFP across many tasks and found it consistently improved performance and speed.
What this means in practice
- •For mobile app developers: Run smaller but capable language models on phones using the KDFP method to keep performance high while saving computational resources.
- •For cloud service engineers: Improve training efficiency and reduce costs when deploying compact language models by applying KDFP knowledge distillation.
Authors
Ryan Swift, Konstantinos Psounis
Abstract
Knowledge distillation is an established technique for improving the capabilities of small, efficient student models by training them with the representations of larger, more capable teacher models. Much of the recent work in the distillation of large language models (LLMs) has focused on distilling abilities learned during post-training, such as instruction following, chain-of-thought reasoning, and tool usage. This has left a large research gap in general knowledge distillation for LLMs, which is essential for developing efficient and private systems suitable for deployment on edge devices. We take a first-principles approach, evaluating previous lessons from prior works and conducting new explorations to develop a distillation methodology suitable for modern LLMs. We present KDFP, a novel methodology for white-box general knowledge distillation in LLMs. We demonstrate that KDFP outperforms existing methods by 1.6% $-$ 4.9% across 9 benchmarks while increasing training efficiency by up to 99.1% through ephemeral parameter reduction.