Multi-Stage Balanced Distillation: Addressing Long-Tail Challenges in Sequence-Level Knowledge Distillation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhou, Yuhang, Zhu, Jing, Xu, Paiheng, Liu, Xiaoyu, Wang, Xiyao, Koutra, Danai, Ai, Wei, Huang, Furong
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917809014964224
author Zhou, Yuhang
Zhu, Jing
Xu, Paiheng
Liu, Xiaoyu
Wang, Xiyao
Koutra, Danai
Ai, Wei
Huang, Furong
author_facet Zhou, Yuhang
Zhu, Jing
Xu, Paiheng
Liu, Xiaoyu
Wang, Xiyao
Koutra, Danai
Ai, Wei
Huang, Furong
contents Large language models (LLMs) have significantly advanced various natural language processing tasks, but deploying them remains computationally expensive. Knowledge distillation (KD) is a promising solution, enabling the transfer of capabilities from larger teacher LLMs to more compact student models. Particularly, sequence-level KD, which distills rationale-based reasoning processes instead of merely final outcomes, shows great potential in enhancing students' reasoning capabilities. However, current methods struggle with sequence level KD under long-tailed data distributions, adversely affecting generalization on sparsely represented domains. We introduce the Multi-Stage Balanced Distillation (BalDistill) framework, which iteratively balances training data within a fixed computational budget. By dynamically selecting representative head domain examples and synthesizing tail domain examples, BalDistill achieves state-of-the-art performance across diverse long-tailed datasets, enhancing both the efficiency and efficacy of the distilled models.
format Preprint
id arxiv_https___arxiv_org_abs_2406_13114
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Multi-Stage Balanced Distillation: Addressing Long-Tail Challenges in Sequence-Level Knowledge Distillation
Zhou, Yuhang
Zhu, Jing
Xu, Paiheng
Liu, Xiaoyu
Wang, Xiyao
Koutra, Danai
Ai, Wei
Huang, Furong
Computation and Language
Artificial Intelligence
Large language models (LLMs) have significantly advanced various natural language processing tasks, but deploying them remains computationally expensive. Knowledge distillation (KD) is a promising solution, enabling the transfer of capabilities from larger teacher LLMs to more compact student models. Particularly, sequence-level KD, which distills rationale-based reasoning processes instead of merely final outcomes, shows great potential in enhancing students' reasoning capabilities. However, current methods struggle with sequence level KD under long-tailed data distributions, adversely affecting generalization on sparsely represented domains. We introduce the Multi-Stage Balanced Distillation (BalDistill) framework, which iteratively balances training data within a fixed computational budget. By dynamically selecting representative head domain examples and synthesizing tail domain examples, BalDistill achieves state-of-the-art performance across diverse long-tailed datasets, enhancing both the efficiency and efficacy of the distilled models.
title Multi-Stage Balanced Distillation: Addressing Long-Tail Challenges in Sequence-Level Knowledge Distillation
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2406.13114