tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
student models on the rationales in the sequence
level to improve their reasoning ability Ho et al
2022 Shridhar et al 2022 Hsieh et al 2023
which proves to be more effective In this work we
mainly discuss using a teacher model to generate
rationales and improve the students reasoning abil
ity on a longtailed dataset Despite the progress
of KD in the LLM era existing works fail to estab
lish a pipeline to gain knowledge from longtailed
datasets with the sequencelevel KD as few ratio
nale examples are provided for tail knowledge
LongTail Learning focuses on longtail dis
tributed data and has been an emerging topic of
interest in the NLP community Liu et al 2019
Wang et al 2017 Godbole and Jia 2022 Dai et al
2023 Zhang et al 2022 Liu et al 2024c Ap
proaches to solving the longtail problem include
rebalancing information augmentation and mod
ule improvement Zhang et al 2021 He et al2021 Zhu et al 2023 Cui et al 2021 Despite
the importance of longtail learning studies have
shown that LLMs struggle to learn longtail knowl
edge Kandpal et al 2023 Sun et al 2023a In
this work we propose to improve LLMs ability to
learn longtail knowledge via multistage distilla
tion over balanced datasets
Active Learning aims to reduce labeling effort
by selecting only the most useful examples Tra
ditional active learning can be categorized into
uncertaintybased methods Prabhu et al 2019
Margatina et al 2021 Wang et al 2023 and
diversitybased methods Ru et al 2020 Ash
et al 2019 In the LLM era active learning has
been used to reduce human annotation costs by
1 strategically selecting the most informative ex
amples for human feedback or annotation Mar
gatina et al 2023 Osband et al 2022 Wang et al
2020 and 2 integrating language models as anno
tators within an active learning framework without
human supervision Xiao et al 2023 Rouzegar
and Makrehchi 2024 Zhang et al 2023 Li et al
2024c In this work we propose to solve the long
tail problem in the student LLM by actively distill
ing knowledge from a blackbox teacher LLM to
meet the budget requirement
3 Methodology
31 Problem Statement
We define our research problem as follows Given
the teacher LLM Mt the student LLM Ms
a longtailed dataset Dwith domain number
d1 d2 d lforldomains in total and a fixed
budget Bto query the teacher we seek to propose
an efficient framework to finetune an effective and
robust student model Ms overD
32 Overall Approach
To mitigate the performance bias in KD caused
by longtailed datasets within budget constraints
we employ a strategy that combines synthetic
data augmentation with active selection This ap
proach ensures effective finetuning across both
wellrepresented head and underrepresented
tail domains As depicted in Figure 1 we pro
pose a multistage framework to create the training
dataiteratively
We operate in a poolbased setting where a large
dataset denoted as D is available but lacks anno
tations from a teacher modelAt each stage of our BalDistill process we first
implement a balancing policy which we have de
signed to determine the appropriate data distribu
tion for each domain within the training batch This
policy is based on the principles of data equality
and training effectiveness across domains aiming
to optimize learning outcomes despite data scarcity
in certain areas The total number of stages is pre
defined based on the consideration of efficiency
and the optimal performance
For domains wellrepresented in our dataset D
referred to as head domains we employ active
selection techniques Touvron et al 2023 Yuan
et al 2020 using the finetuned student model
Msto identify and extract the most informative
examples from the pool Conversely for domains
lacking sufficient data tail domains we utilize
the teacher model Mtto generate both synthetic
samples and corresponding annotations enriching
the training material available
After selecting andor generating these samples
we query the teacher model to provide detailed
rationales for examples in the training batch These
annotated examples are then used to finetune the
student model Msin preparation for the next stage
Detailed descriptions of these components along
with the algorithms outlining this procedure are
presented in Algorithm 1 in Appendix B
33 Balancing Policy
Considering Ktotal stages in our framework we
first evenly divide our budget BintoKparts which
means that for each stage we create a small train
ing batch withB
Kexamples extracted from Dwith
teacherannotated rationales Within a small train

student models on the rationales in the sequence

level to improve their reasoning ability Ho et al

2022 Shridhar et al 2022 Hsieh et al 2023

which proves to be more effective In this work we

mainly discuss using a teacher model to generate

rationales and improve the students reasoning abil

ity on a longtailed dataset Despite the progress

of KD in the LLM era existing works fail to estab

lish a pipeline to gain knowledge from longtailed

datasets with the sequencelevel KD as few ratio

nale examples are provided for tail knowledge

LongTail Learning focuses on longtail dis

tributed data and has been an emerging topic of

interest in the NLP community Liu et al 2019

Wang et al 2017 Godbole and Jia 2022 Dai et al

2023 Zhang et al 2022 Liu et al 2024c Ap

proaches to solving the longtail problem include

rebalancing information augmentation and mod

ule improvement Zhang et al 2021 He et al2021 Zhu et al 2023 Cui et al 2021 Despite

the importance of longtail learning studies have

shown that LLMs struggle to learn longtail knowl

edge Kandpal et al 2023 Sun et al 2023a In

this work we propose to improve LLMs ability to

learn longtail knowledge via multistage distilla

tion over balanced datasets

Active Learning aims to reduce labeling effort

by selecting only the most useful examples Tra

ditional active learning can be categorized into

uncertaintybased methods Prabhu et al 2019

Margatina et al 2021 Wang et al 2023 and

diversitybased methods Ru et al 2020 Ash

et al 2019 In the LLM era active learning has

been used to reduce human annotation costs by

1 strategically selecting the most informative ex

amples for human feedback or annotation Mar

gatina et al 2023 Osband et al 2022 Wang et al

2020 and 2 integrating language models as anno

tators within an active learning framework without

human supervision Xiao et al 2023 Rouzegar

and Makrehchi 2024 Zhang et al 2023 Li et al

2024c In this work we propose to solve the long

tail problem in the student LLM by actively distill

ing knowledge from a blackbox teacher LLM to

meet the budget requirement

3 Methodology

31 Problem Statement

We define our research problem as follows Given

the teacher LLM Mt the student LLM Ms

a longtailed dataset Dwith domain number

d1 d2 d lforldomains in total and a fixed

budget Bto query the teacher we seek to propose

an efficient framework to finetune an effective and

robust student model Ms overD

32 Overall Approach

To mitigate the performance bias in KD caused

by longtailed datasets within budget constraints

we employ a strategy that combines synthetic

data augmentation with active selection This ap

proach ensures effective finetuning across both

wellrepresented head and underrepresented

tail domains As depicted in Figure 1 we pro

pose a multistage framework to create the training

dataiteratively

We operate in a poolbased setting where a large

dataset denoted as D is available but lacks anno

tations from a teacher modelAt each stage of our BalDistill process we first

implement a balancing policy which we have de

signed to determine the appropriate data distribu

tion for each domain within the training batch This

policy is based on the principles of data equality

and training effectiveness across domains aiming

to optimize learning outcomes despite data scarcity

in certain areas The total number of stages is pre

defined based on the consideration of efficiency

and the optimal performance

For domains wellrepresented in our dataset D

referred to as head domains we employ active

selection techniques Touvron et al 2023 Yuan

et al 2020 using the finetuned student model

Msto identify and extract the most informative

examples from the pool Conversely for domains

lacking sufficient data tail domains we utilize

the teacher model Mtto generate both synthetic

samples and corresponding annotations enriching

the training material available

After selecting andor generating these samples

we query the teacher model to provide detailed

rationales for examples in the training batch These

annotated examples are then used to finetune the

student model Msin preparation for the next stage

Detailed descriptions of these components along

with the algorithms outlining this procedure are

presented in Algorithm 1 in Appendix B

33 Balancing Policy

Considering Ktotal stages in our framework we

first evenly divide our budget BintoKparts which

means that for each stage we create a small train

ing batch withB

Kexamples extracted from Dwith

teacherannotated rationales Within a small train