Sakana AI· Sakana modelsResearch paper

TAID: A Novel Method for Efficient Knowledge Transfer from Large Language Models to Small Language Models

Sakana AI introduces TAID, a novel knowledge distillation method for transferring knowledge from large language models to small ones. They release two models: TAID-LLM-1.5B (English) and TinySwallow-1.5B (Japanese), with TinySwallow-1.5B achieving state-of-the-art performance among similar-sized models. The research is accepted as a Spotlight paper at ICLR 2025.

From the source

Key highlights of this release: We introduce TAID (Temporally Adaptive Interpolated Distillation), a novel knowledge distillation method that enables efficient transfer of knowledge from large language models to smaller ones. In collaboration with the Institute of Science Tokyo, we have developed TinySwallow-1.5B, a compact Japanese language model that enables offline execution on edge devices like smartphones. Our research has been accepted as a Spotlight paper at ICLR, a top international conference in machine learning.

sakana.ai

Announced

24 Feb 2025 15:00 UTC