In-Context Learning Distillation for Efficient Few-Shot Fine-Tuning

TL;DR

This paper introduces a knowledge distillation method for in-context learning that significantly reduces model size and memory usage while improving out-of-domain accuracy in natural language inference tasks.

Contribution

It presents a novel in-context learning distillation technique that enhances model efficiency and performance beyond traditional prompt-based fine-tuning methods.

Findings

Model size reduced from 2.5GB to 0.25GB.

Out-of-domain accuracy improved by nearly 50%.

Memory consumption decreased by up to 60%.

Abstract

We applied few-shot in-context learning on the OPT-1.3B model for the natural language inference task and employed knowledge distillation to internalize the context information, reducing model parameter from 1.3B to 125M and achieving a size reduction from 2.5GB to 0.25GB. Compared to using in-context learning alone on similarly sized models, this context distillation approach achieved a nearly 50% improvement in out-of-domain accuracy, demonstrating superior knowledge transfer capabilities over prompt-based methods. Furthermore, this approach reduced memory consumption by up to 60% while delivering a 20% improvement in out-of-domain accuracy compared to conventional pattern-based fine-tuning.