tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
ficiency to compensate for data loss in the head
domain Note that for Math dataset BalDistill can
only achieve comparable results with the baseline
methods on the last two tail domains precalculus
and probability and we conjecture that the high
difficulty in these two domains prevents the teacher
from composing highquality synthetic data
52 Ablation Study
After showing the superiority of our overall frame
work our next step is to verify the effectiveness
of each component in the proposed method We
compare our framework with the ablated methods
1Balance Finetune CoT We adopt a naive balMethodR52 Reuters AbstractiveQA MultichoiceQA Math
macrof1 microf1 macrof1 microf1 macrof1 microf1 macroacc microacc macroacc microacc
Zeroshot CoT 089 230 074 161 760 759 2467 2495 757 868
Random Finetune 4595 9144 2801 7468 3762 3721 6123 5596 1012 948
Random Finetune CoT 5970 8946 2735 7053 5257 5288 7609 7412 1662 1520
Duplicate Finetune CoT 4656 7179 2676 6284 5132 5137 7592 7399 1698 1505
BalDistill N 5962 8249 2809 6240 5270 5292 7660 7343 1790 1634
BalDistill A 5893 8747 3295 6977 5320 5290 7717 7473 1866 1742
Table 2 Performance of proposed BalDistill framework and other baselines across five longtailed datasets
The best performance is marked in bold The performance of finetuned student models with our framework can
outperform other baselines in macroaverages on multiple longtailed datasets
Method R52 Reuters AbsQA MCQA Math
Budget Setting 1
Random FT CoT 5970 2735 5257 7609 1520
Balance FT CoT 5147 2712 5222 7598 1629
Active FT CoT 5949 2975 5314 7664 1561
BalDistill N 5962 2809 5270 7660 1634
BalDistill A 5893 3295 5320 7717 1742
Budget Setting 2
Random FT CoT 6488 3342 5371 7292 1519
Balance FT CoT 6055 3279 5029 7629 1573
Active FT CoT 6454 3133 5305 7626 1591
BalDistill N 5935 3276 5386 7617 1759
BalDistill A 6584 3277 5349 7711 1759
Table 3 Effects of active learning and adaptive bal
ancing in BalDistill framework Results of finetuned
student models on five datasets outperform methods
with only balancing Balance FT CoT with only active
learning Active FT CoT
ancing policy to construct the training set and query
the teacher model to compose inputs in the tail do
mains 2 Active Finetune CoT We only keep the
active learning component but remove the data aug
mentation part The experiment setting is similar
to the setup in Section 51 and we present the per
formance of each method with two budget settings
in Table 3
Both active selection and adaptive balancing
bring salient performance boost From Table 3
we find that our BalDistill A method obtains the
best performance in 710 comparison cases which
demonstrates the effectiveness of each framework
component We notice that by simply adding the
active learning strategy Random Finetune CoT vs
Active Finetune CoT the finetuned student model
can achieve a performance boost in most cases
with an average relative improvement of 143
This observation is consistent with the findings in
previous work for Bert models Devlin et al 2019
on the longtailed data Dor et al 2020
However when we add data augmentation from
the teacher with the naive balancing policy Bal
Head T ail048050052054056058F1a AbstractiveQA
Head T ail020030040050F1 b Reuters
Head T ail065070075080085Acc
c MultichoiceQA
Head T ail010015020025AccRandom
Balance
Active
BalDistill N
BalDistill A d Math
Figure 3 Performance of proposed method BalDis
till and ablated methods on head and tail domains
BalDistill A can achieve better results on head do
mains and outperform the Active FT CoT method on
tail domains which demonstrates the effectiveness of
each component in our BalDistill A framework
ance Finetune CoT vs Random Finetune CoT
BalDistill N vs Active Finetune CoT this oper
ation does not substantially improve performance
This finding suggests the superiority of our adap
tive balancing policy
To probe the detailed reasons for the result pat
terns above we visualize the macroaverage perfor
mance of these methods on inputs from head and
tail domains in Figure 3 The splitting criteria for
each dataset can be found in Appendix A We find
that for methods with naive balancing policy Bal
ance Finetune CoT and BalDistill N there exists
a significant performance drop on head domains
due to filtering a large proportion of data and our
method with adaptive balancing can achieve com
parable performance on head domains The obser
vation suggests the effectiveness of active selection
for head domains and the importance of adaptive
balancing for the finetuned student to select the
uncertain ones preciselyFor performance in tail domains our proposed
method with adaptive balancing and teacher aug

ficiency to compensate for data loss in the head

domain Note that for Math dataset BalDistill can

only achieve comparable results with the baseline

methods on the last two tail domains precalculus

and probability and we conjecture that the high

difficulty in these two domains prevents the teacher

from composing highquality synthetic data

52 Ablation Study

After showing the superiority of our overall frame

work our next step is to verify the effectiveness

of each component in the proposed method We

compare our framework with the ablated methods

1Balance Finetune CoT We adopt a naive balMethodR52 Reuters AbstractiveQA MultichoiceQA Math

macrof1 microf1 macrof1 microf1 macrof1 microf1 macroacc microacc macroacc microacc

Zeroshot CoT 089 230 074 161 760 759 2467 2495 757 868

Random Finetune 4595 9144 2801 7468 3762 3721 6123 5596 1012 948

Random Finetune CoT 5970 8946 2735 7053 5257 5288 7609 7412 1662 1520

Duplicate Finetune CoT 4656 7179 2676 6284 5132 5137 7592 7399 1698 1505

BalDistill N 5962 8249 2809 6240 5270 5292 7660 7343 1790 1634

BalDistill A 5893 8747 3295 6977 5320 5290 7717 7473 1866 1742

Table 2 Performance of proposed BalDistill framework and other baselines across five longtailed datasets

The best performance is marked in bold The performance of finetuned student models with our framework can

outperform other baselines in macroaverages on multiple longtailed datasets

Method R52 Reuters AbsQA MCQA Math

Budget Setting 1

Random FT CoT 5970 2735 5257 7609 1520

Balance FT CoT 5147 2712 5222 7598 1629

Active FT CoT 5949 2975 5314 7664 1561

BalDistill N 5962 2809 5270 7660 1634

BalDistill A 5893 3295 5320 7717 1742

Budget Setting 2

Random FT CoT 6488 3342 5371 7292 1519

Balance FT CoT 6055 3279 5029 7629 1573

Active FT CoT 6454 3133 5305 7626 1591

BalDistill N 5935 3276 5386 7617 1759

BalDistill A 6584 3277 5349 7711 1759

Table 3 Effects of active learning and adaptive bal

ancing in BalDistill framework Results of finetuned

student models on five datasets outperform methods

with only balancing Balance FT CoT with only active

learning Active FT CoT

ancing policy to construct the training set and query

the teacher model to compose inputs in the tail do

mains 2 Active Finetune CoT We only keep the

active learning component but remove the data aug

mentation part The experiment setting is similar

to the setup in Section 51 and we present the per

formance of each method with two budget settings

in Table 3

Both active selection and adaptive balancing

bring salient performance boost From Table 3

we find that our BalDistill A method obtains the

best performance in 710 comparison cases which

demonstrates the effectiveness of each framework

component We notice that by simply adding the

active learning strategy Random Finetune CoT vs

Active Finetune CoT the finetuned student model

can achieve a performance boost in most cases

with an average relative improvement of 143

This observation is consistent with the findings in

previous work for Bert models Devlin et al 2019

on the longtailed data Dor et al 2020

However when we add data augmentation from

the teacher with the naive balancing policy Bal

Head T ail048050052054056058F1a AbstractiveQA

Head T ail020030040050F1 b Reuters

Head T ail065070075080085Acc

c MultichoiceQA

Head T ail010015020025AccRandom

Balance

Active

BalDistill N

BalDistill A d Math

Figure 3 Performance of proposed method BalDis

till and ablated methods on head and tail domains

BalDistill A can achieve better results on head do

mains and outperform the Active FT CoT method on

tail domains which demonstrates the effectiveness of

each component in our BalDistill A framework

ance Finetune CoT vs Random Finetune CoT

BalDistill N vs Active Finetune CoT this oper

ation does not substantially improve performance

This finding suggests the superiority of our adap

tive balancing policy

To probe the detailed reasons for the result pat

terns above we visualize the macroaverage perfor

mance of these methods on inputs from head and

tail domains in Figure 3 The splitting criteria for

each dataset can be found in Appendix A We find

that for methods with naive balancing policy Bal

ance Finetune CoT and BalDistill N there exists

a significant performance drop on head domains

due to filtering a large proportion of data and our

method with adaptive balancing can achieve com

parable performance on head domains The obser

vation suggests the effectiveness of active selection

for head domains and the importance of adaptive

balancing for the finetuned student to select the

uncertain ones preciselyFor performance in tail domains our proposed

method with adaptive balancing and teacher aug