Evolving Knowledge Distillation with Large Language Models and Active Learning

TL;DR

EvoKD introduces an active learning-based approach to knowledge distillation from large language models, iteratively improving small models' performance by analyzing weaknesses and generating targeted, challenging training samples.

Contribution

The paper presents a novel active learning framework for knowledge distillation that actively analyzes student weaknesses and guides LLMs to generate more effective training data.

Findings

EvoKD improves small model performance on NLP tasks

Active analysis leads to more diverse and challenging training samples

Method outperforms traditional distillation approaches

Abstract

Large language models (LLMs) have demonstrated remarkable capabilities across various NLP tasks. However, their computational costs are prohibitively high. To address this issue, previous research has attempted to distill the knowledge of LLMs into smaller models by generating annotated data. Nonetheless, these works have mainly focused on the direct use of LLMs for text generation and labeling, without fully exploring their potential to comprehend the target task and acquire valuable knowledge. In this paper, we propose EvoKD: Evolving Knowledge Distillation, which leverages the concept of active learning to interactively enhance the process of data generation using large language models, simultaneously improving the task capabilities of small domain model (student model). Different from previous work, we actively analyze the student model's weaknesses, and then synthesize labeled samples based on the analysis. In addition, we provide iterative feedback to the LLMs regarding the student model's performance to continuously construct diversified and challenging samples. Experiments and analysis on different NLP tasks, namely, text classification and named entity recognition show the effectiveness of EvoKD.