tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
uncertainty level as perceived by the student model
IFD is calculated as the ratio of the perplexity of
generating a response ywith an input xto the per
plexity of generating ywithout xIFDx y
PPLyx
PPLy where PPL represents perplexity a metric
widely used to evaluate language model perfor
mance Jelinek et al 1977 Studies have shown
that IFD scores offer greater efficiency in data se
lection compared to methods like Kmeans diver
sity or sole reliance on perplexity Li et al 2024a
Settles 2009 Yuan et al 2020
A higher IFD score indicates an increased dif
ficulty for the model in generating the response
highlighting the instances value for training Li
et al 2024a
Unlike the approach in Li et al 2024a which
utilizes groundtruth or advanced LLMgenerated
responses y our setting imposes budge constraints
that prevent such usage Instead we calculate IFDusing rationals ˆysgenerated by the previously fine
tuned student model allowing us to assess the
models selfuncertainty and conserve the anno
tation budget from the teacher model
At last we rank the inputs by their IFD scores
selecting those with the highest values to include
in the bath as specified by the balancing policy
36 Reasoning Generation and Finetuning
Building on methodologies from prior research that
focus on distilling reasoning abilities from black
box LLMs Ho et al 2022 Hsieh et al 2023
we employ a zeroshot CoT approach where the
teacher model is prompted to generate a reasoning
explanation ˆytfor the samples in our constructed
training batch This zeroshot setting is crucial for
demonstrating the models ability to reason based
on its preexisting knowledge alone Brown et al
2020 In our experimental setup which utilizes
labeled datasets lacking rationale annotations the
final ground truth answer is included in the prompt
This inclusion ensures that the generated explana
tions are aligned with the correct outcomes en
hancing the accuracy and relevance of the CoT
reasoning It is important to note that for synthetic
samples generated from tail domains in 34 we do
not perform additional annotations in this part to
maintain adherence to budget constraints
After gathering the required samples and their
associated rationales in the training batch we in
tegrate this batch with the annotated data accumu
lated from previous stages This approach ensures
that our student model is exposed to a diverse and
comprehensive dataset which helps mitigate the
risk of overfitting a common challenge in ma
chine learning models as identified in prior studies
Dor et al 2020 To facilitate this we reinitialize
and finetune the student model on the compiled
rationale sequences from scratch at each stage
The finetuning is performed using autoregres
sive language modeling with a crossentropy loss
aligning with the original pretraining objectives of
the student model Touvron et al 2023
4 Experiment
Through our extensive empirical analysis we aim
to address the following research questions
RQ1 How effective is our KD framework com
pared to previous KD baseline methods
RQ2 How important is each component balanc
ing policy and active learning to the frameworkRQ3 How well does our method perform with
different student models and budget restrictions
Dataset To verify the effectiveness of our frame
work on various reasoning tasks we evaluate our
method on five longtailed datasets following pre
vious work Yu et al 2023 Dai et al 2023 Huang
et al 2021 These include text classification
R52 and Reuters Hayes and Weinstein 1990
question answering AbstractiveQA and Multiple
choiceQA Dai et al 2023 and arithmetic MATH
Hendrycks et al 2021 For text classification
datasets we treat the label of inputs as the do
main for other datasets the domain information
of inputs is annotated as metadata from the data
provider The detailed construction process and do
main information for these datasets can be found in
Appendix A We also show two example distribu
tions of the datasets in Figure 5 in Appendix A For
each dataset we prepare two budget settings for
the experiment In Table 1 we present the budget
number the test number the domain number and
the evaluation metric of all five datasets
Evaluation metrics Since we are dealing with
longtailed imbalanced data for each dataset we
choose to use both the micro and macroaverages
to evaluate the method robustness Henning et al
2022 For the classification datasets R52 and
Reuters we report microF1 and macroF1 where
microF1 is a global average F1 score and macro
F1 is computed by taking the unweighted mean
of all the perclass F1 scores Harbecke et al
2022 For other datasets we also report the micro
macroF1 for AbstractiveQA datasets and micro

uncertainty level as perceived by the student model

IFD is calculated as the ratio of the perplexity of

generating a response ywith an input xto the per

plexity of generating ywithout xIFDx y

PPLyx

PPLy where PPL represents perplexity a metric

widely used to evaluate language model perfor

mance Jelinek et al 1977 Studies have shown

that IFD scores offer greater efficiency in data se

lection compared to methods like Kmeans diver

sity or sole reliance on perplexity Li et al 2024a

Settles 2009 Yuan et al 2020

A higher IFD score indicates an increased dif

ficulty for the model in generating the response

highlighting the instances value for training Li

et al 2024a

Unlike the approach in Li et al 2024a which

utilizes groundtruth or advanced LLMgenerated

responses y our setting imposes budge constraints

that prevent such usage Instead we calculate IFDusing rationals ˆysgenerated by the previously fine

tuned student model allowing us to assess the

models selfuncertainty and conserve the anno

tation budget from the teacher model

At last we rank the inputs by their IFD scores

selecting those with the highest values to include

in the bath as specified by the balancing policy

36 Reasoning Generation and Finetuning

Building on methodologies from prior research that

focus on distilling reasoning abilities from black

box LLMs Ho et al 2022 Hsieh et al 2023

we employ a zeroshot CoT approach where the

teacher model is prompted to generate a reasoning

explanation ˆytfor the samples in our constructed

training batch This zeroshot setting is crucial for

demonstrating the models ability to reason based

on its preexisting knowledge alone Brown et al

2020 In our experimental setup which utilizes

labeled datasets lacking rationale annotations the

final ground truth answer is included in the prompt

This inclusion ensures that the generated explana

tions are aligned with the correct outcomes en

hancing the accuracy and relevance of the CoT

reasoning It is important to note that for synthetic

samples generated from tail domains in 34 we do

not perform additional annotations in this part to

maintain adherence to budget constraints

After gathering the required samples and their

associated rationales in the training batch we in

tegrate this batch with the annotated data accumu

lated from previous stages This approach ensures

that our student model is exposed to a diverse and

comprehensive dataset which helps mitigate the

risk of overfitting a common challenge in ma

chine learning models as identified in prior studies

Dor et al 2020 To facilitate this we reinitialize

and finetune the student model on the compiled

rationale sequences from scratch at each stage

The finetuning is performed using autoregres

sive language modeling with a crossentropy loss

aligning with the original pretraining objectives of

the student model Touvron et al 2023

4 Experiment

Through our extensive empirical analysis we aim

to address the following research questions

RQ1 How effective is our KD framework com

pared to previous KD baseline methods

RQ2 How important is each component balanc

ing policy and active learning to the frameworkRQ3 How well does our method perform with

different student models and budget restrictions

Dataset To verify the effectiveness of our frame

work on various reasoning tasks we evaluate our

method on five longtailed datasets following pre

vious work Yu et al 2023 Dai et al 2023 Huang

et al 2021 These include text classification

R52 and Reuters Hayes and Weinstein 1990

question answering AbstractiveQA and Multiple

choiceQA Dai et al 2023 and arithmetic MATH

Hendrycks et al 2021 For text classification

datasets we treat the label of inputs as the do

main for other datasets the domain information

of inputs is annotated as metadata from the data

provider The detailed construction process and do

main information for these datasets can be found in

Appendix A We also show two example distribu

tions of the datasets in Figure 5 in Appendix A For

each dataset we prepare two budget settings for

the experiment In Table 1 we present the budget

number the test number the domain number and

the evaluation metric of all five datasets

Evaluation metrics Since we are dealing with

longtailed imbalanced data for each dataset we

choose to use both the micro and macroaverages

to evaluate the method robustness Henning et al

2022 For the classification datasets R52 and

Reuters we report microF1 and macroF1 where

microF1 is a global average F1 score and macro

F1 is computed by taking the unweighted mean

of all the perclass F1 scores Harbecke et al

2022 For other datasets we also report the micro

macroF1 for AbstractiveQA datasets and micro