tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
1 Introduction
Large language models LLMs like GPT4 and
LLaMA have revolutionized tasks ranging from
text generation to language translation through their
deep understanding and generation of humanlike
text OpenAI 2023 Touvron et al 2023 Chiang
et al 2023 Jiang et al 2023 Despite their suc
cess the deployment of these models is hindered by
their substantial size and computational demands
especially in environments with limited resources
Knowledge distillation KD offers a viable so
lution by transferring knowledge from expensive
teacher models to smaller efficient student models
Specifically sequencelevel KD focuses on distill
ing rationalebased reasoning processes rather than
Equal contributionfinal outcomes It leverages the teachers reasoning
processes encapsulated in chainofthought CoT
rationales to enhance the student models gener
ative capabilities Kim and Rush 2016 Ho et al
2022 Shridhar et al 2022 Hsieh et al 2023
However there are a few challenges to fully
leverage the power of sequencelevel KD as fol
lows C1 Sequencelevel KD encounters signifi
cant challenges when training with longtailed data
distributions which are prevalent in realworld sce
narios data often follows a powerlaw distribu
tion with a few dominant classes head and many
rare classes tail Liu et al 2019 Such distri
butions feature a few dominant classes and many
underrepresented ones leading to models that gen
eralize poorly on sparsely represented domains
C2 Traditional KD methods in the text area to
solve longtail challenges often reliant on direct
access to model weights or loss adjustment primar
ily suited for straightforward classification tasks
Zhou et al 2023 Schick and Schütze 2021 Dai
et al 2023 Zhang et al 2022 Tepper et al 2020
falter under the complexities of sequencelevel KD
especially when the teacher model is a black box
and the task is generative which is our target C3
Addressing this imbalance is critical yet resource
intensive as it typically requires generating a large
volume of synthetic data to balance the dataset Tep
per et al 2020 Moreover naively upsampling
the longtailed dataset may dramatically increase
the number of calls to the teacher models Budget
constraints play a crucial role in KD for blackbox
LLMs as querying the teacher for rationales can
be costly and timeconsuming Chen et al 2023
Zhou and Ai 2024
Our proposed solution the MultiStage Bal
anced Distillation BalDistill tackles all the chal
lenges above by strategically generating balanced
training sets within budget constraints and itera
tively finetuning the student model with actively
selected and synthetic data for multiple stagesarXiv240613114v1 csCL 19 Jun 2024Figure 1 Overview of the proposed iterative BalDistill framework The framework is composed of multiple
stages For each stage we apply the balancing policy to decide the data distribution in the training batch For head
domains with sufficient data we actively extract the examples by IFD metrics using the student model For the
tail domains we call the teacher model to generate the synthetic examples and the corresponding rationales The
teacher model finally annotates the balanced training batch and finetunes the student model
BalDistill progressively refines the training data
by selecting key examples from wellrepresented
domains and generating necessary synthetic data
for underrepresented ones ensuring comprehen
sive domain coverage and model robustness By
dynamically selecting representative head domain
examples and synthesizing tail domain examples
BalDistill achieves stateoftheart SoTA perfor
mance on various longtailed datasets enhancing
both the efficiency and efficacy of the method
Our contributions are summarized as follows
Innovative Problem Framing We address the
underexplored challenge of applying sequence
level KD to longtailed distributions where the
teacher model is a blackbox LLM
Strategic Framework BalDistill innovatively
combines active example selection with synthetic
data generation for multiple stages to maintain
training balance within predefined budget limits
SoTA Performance Our framework demonstra
bly improves the student models effectiveness
and robustness across diverse domains setting
new benchmarks in performance We empiri
cally demonstrate that our distilled student mod
els achieve stateoftheart performance across a
range of benchmark datasets
2 Related Work
Knowledge Distillation uses the outputs of a larger
LLM Teacher such as ChatGPT OpenAI 2023to train a smaller model Student such as LLaMa
7B Touvron et al 2023 For details of knowl
edge distillation KD of large language models
we refer to the survey for more details Xu et al
2024 In this work we focus on KD with black
box teacher models There are two lines of work
with respect to knowledge distillation The first
is to ask teacher models to generate the final an
swers and to finetune on the final answers Zhou
et al 2023 Schick and Schütze 2021 Another
line of work asks teacher models to generate ra
tionales at the reasoning process and finetunes

1 Introduction

Large language models LLMs like GPT4 and

LLaMA have revolutionized tasks ranging from

text generation to language translation through their

deep understanding and generation of humanlike

text OpenAI 2023 Touvron et al 2023 Chiang

et al 2023 Jiang et al 2023 Despite their suc

cess the deployment of these models is hindered by

their substantial size and computational demands

especially in environments with limited resources

Knowledge distillation KD offers a viable so

lution by transferring knowledge from expensive

teacher models to smaller efficient student models

Specifically sequencelevel KD focuses on distill

ing rationalebased reasoning processes rather than

Equal contributionfinal outcomes It leverages the teachers reasoning

processes encapsulated in chainofthought CoT

rationales to enhance the student models gener

ative capabilities Kim and Rush 2016 Ho et al

2022 Shridhar et al 2022 Hsieh et al 2023

However there are a few challenges to fully

leverage the power of sequencelevel KD as fol

lows C1 Sequencelevel KD encounters signifi

cant challenges when training with longtailed data

distributions which are prevalent in realworld sce

narios data often follows a powerlaw distribu

tion with a few dominant classes head and many

rare classes tail Liu et al 2019 Such distri

butions feature a few dominant classes and many

underrepresented ones leading to models that gen

eralize poorly on sparsely represented domains

C2 Traditional KD methods in the text area to

solve longtail challenges often reliant on direct

access to model weights or loss adjustment primar

ily suited for straightforward classification tasks

Zhou et al 2023 Schick and Schütze 2021 Dai

et al 2023 Zhang et al 2022 Tepper et al 2020

falter under the complexities of sequencelevel KD

especially when the teacher model is a black box

and the task is generative which is our target C3

Addressing this imbalance is critical yet resource

intensive as it typically requires generating a large

volume of synthetic data to balance the dataset Tep

per et al 2020 Moreover naively upsampling

the longtailed dataset may dramatically increase

the number of calls to the teacher models Budget

constraints play a crucial role in KD for blackbox

LLMs as querying the teacher for rationales can

be costly and timeconsuming Chen et al 2023

Zhou and Ai 2024

Our proposed solution the MultiStage Bal

anced Distillation BalDistill tackles all the chal

lenges above by strategically generating balanced

training sets within budget constraints and itera

tively finetuning the student model with actively

selected and synthetic data for multiple stagesarXiv240613114v1 csCL 19 Jun 2024Figure 1 Overview of the proposed iterative BalDistill framework The framework is composed of multiple

stages For each stage we apply the balancing policy to decide the data distribution in the training batch For head

domains with sufficient data we actively extract the examples by IFD metrics using the student model For the

tail domains we call the teacher model to generate the synthetic examples and the corresponding rationales The

teacher model finally annotates the balanced training batch and finetunes the student model

BalDistill progressively refines the training data

by selecting key examples from wellrepresented

domains and generating necessary synthetic data

for underrepresented ones ensuring comprehen

sive domain coverage and model robustness By

dynamically selecting representative head domain

examples and synthesizing tail domain examples

BalDistill achieves stateoftheart SoTA perfor

mance on various longtailed datasets enhancing

both the efficiency and efficacy of the method

Our contributions are summarized as follows

Innovative Problem Framing We address the

underexplored challenge of applying sequence

level KD to longtailed distributions where the

teacher model is a blackbox LLM

Strategic Framework BalDistill innovatively

combines active example selection with synthetic

data generation for multiple stages to maintain

training balance within predefined budget limits

SoTA Performance Our framework demonstra

bly improves the student models effectiveness

and robustness across diverse domains setting

new benchmarks in performance We empiri

cally demonstrate that our distilled student mod

els achieve stateoftheart performance across a

range of benchmark datasets

2 Related Work

Knowledge Distillation uses the outputs of a larger

LLM Teacher such as ChatGPT OpenAI 2023to train a smaller model Student such as LLaMa

7B Touvron et al 2023 For details of knowl

edge distillation KD of large language models

we refer to the survey for more details Xu et al

2024 In this work we focus on KD with black

box teacher models There are two lines of work

with respect to knowledge distillation The first

is to ask teacher models to generate the final an

swers and to finetune on the final answers Zhou

et al 2023 Schick and Schütze 2021 Another

line of work asks teacher models to generate ra

tionales at the reasoning process and finetunes