tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
macroaccuracy for Math and MultichoiceQA
datasets Note that the F1 score for the AbstractQA
is the wordlevel F1 score between the token list
of ground truth answer and the generated answer
different from the F1 for the classification task
Model setup For the teacher model we use GPT
4 OpenAI 2023 to generate the CoT rationales for
each dataset We choose between Llama27B and
Llama38B as our student models Touvron et al
2023 We include the detailed configurations and
implementations of the model in Appendix B
Baseline methods We experiment with two vari
ants of our proposed method with different balanc
ing policies as discussed in Section 3 In our first
framework BalDistill N we use naive balanc
ing policy and for second framework BalDistillDataset Budget Test Domain Metric Task
R52 26005200 2570 52 F1 TC
Reuters 45009000 3745 90 F1 TC
Abstractive QA 500010000 10000 5 F1 QA
Multichoice QA 50008000 10520 10 Accuracy QA
Math 21003500 5000 7 Accuracy Arithmetic
Table 1 Dataset statistics TC and QA represent the
text classification and question answering respectively
A we leverage adaptive balancing We compare
our framework with multiple baseline methods 1
Zeroshot CoT We directly prompt the student
model to infer on the test data Kojima et al 2022
2Random Finetune We randomly collect sam
ples from the training data until the budget con
straint is met and finetune student models on the
final groundtruth labels Radford et al 2019 3
Random FinetuneCoT We randomly collect and
use CoT rationales from the teacher model for stu
dent finetuning Ho et al 2022 Yao et al 2022
He et al 2023 4 Duplicate FinetuneCoT We
construct the training data with a naive balancing
policy and for tail domains we duplicate the inputs
to satisfy the policy requirement
5 Results
51 Comparison with Baseline Methods
BalDistill framework outperforms Random
Finetune and Duplicate Finetune methods
We use Llama3 as the student model GPT4 as
the teacher model and choose the smaller budget
for each dataset in Table 1 as our experiment set
tings for this subsection We present the overall
macro and microaverage results of the proposed
frameworks and the baseline methods in Table 2
From Table 2 we first observe that on the long
tailed dataset the methods finetuned on teacher
generated rationales CoT can significantly out
perform the groundtruth finetuning method Ran
dom Finetune which emphasizes the necessity of
teachergenerated reasoning steps in the KD
Among all sequencelevel KD methods our pro
posed BalDistill N and BalDistill A achieve the
best average performance across various datasets
on macroaverages which obtain an average rel
ative improvement of 224 and 681 respec
tively compared to the Random Finetune CoT base
line The performance boost in BalDistill N im
plies the effectiveness of replacing the naive bal
ancing policy with adaptive balancing
Moreover we note that the Duplicate Finetune
CoT baseline fails to compete with the Random
Finetune CoT method in most cases which indi
343 289 224 108 35
Domain Proportion 045048050053055058F1a AbstractiveQA
294 09 05 03 02 01
Domain Proportion 000020040060080100F1 b Reuters
1861751551201168854483028
Domain Proportion 060070080090Acc
c MultichoiceQA
237 181 174 109 108 96 95
Domain Proportion 010020030AccRandom
Duplicate
BalDistill A d Math
Figure 2 Performance of proposed method and base
lines on different domains Xaxis represents the pro
portion of each domain ranked from head to tail do
mains Our proposed BalDistill method can achieve
comparable results on head domains and outperform the
baseline method on the tail domains
cates that simply duplicating the input from the tail
domains to ensure balanced data cannot address
the underlying imbalanced data complexity
To perform a detailed analysis of our framework
we visualize the F1 or accuracy score for each do
main of the BalDistill N method and two baseline
methods Random Finetune CoT and Duplicate
Finetune CoT in Figure 2 with the xaxis repre
senting the proportion of each domain in the dataset
in descending order From Figure 2 our proposed
method can achieve comparable results in the head
domains left side of the figure but substantially
outperform the baseline methods in the tail domains
right side of the figure This observation verifies
our expectation in Section 3 where the balancing
policy increases performance in the tail domain
and the active learning part improves the data ef

macroaccuracy for Math and MultichoiceQA

datasets Note that the F1 score for the AbstractQA

is the wordlevel F1 score between the token list

of ground truth answer and the generated answer

different from the F1 for the classification task

Model setup For the teacher model we use GPT

4 OpenAI 2023 to generate the CoT rationales for

each dataset We choose between Llama27B and

Llama38B as our student models Touvron et al

2023 We include the detailed configurations and

implementations of the model in Appendix B

Baseline methods We experiment with two vari

ants of our proposed method with different balanc

ing policies as discussed in Section 3 In our first

framework BalDistill N we use naive balanc

ing policy and for second framework BalDistillDataset Budget Test Domain Metric Task

R52 26005200 2570 52 F1 TC

Reuters 45009000 3745 90 F1 TC

Abstractive QA 500010000 10000 5 F1 QA

Multichoice QA 50008000 10520 10 Accuracy QA

Math 21003500 5000 7 Accuracy Arithmetic

Table 1 Dataset statistics TC and QA represent the

text classification and question answering respectively

A we leverage adaptive balancing We compare

our framework with multiple baseline methods 1

Zeroshot CoT We directly prompt the student

model to infer on the test data Kojima et al 2022

2Random Finetune We randomly collect sam

ples from the training data until the budget con

straint is met and finetune student models on the

final groundtruth labels Radford et al 2019 3

Random FinetuneCoT We randomly collect and

use CoT rationales from the teacher model for stu

dent finetuning Ho et al 2022 Yao et al 2022

He et al 2023 4 Duplicate FinetuneCoT We

construct the training data with a naive balancing

policy and for tail domains we duplicate the inputs

to satisfy the policy requirement

5 Results

51 Comparison with Baseline Methods

BalDistill framework outperforms Random

Finetune and Duplicate Finetune methods

We use Llama3 as the student model GPT4 as

the teacher model and choose the smaller budget

for each dataset in Table 1 as our experiment set

tings for this subsection We present the overall

macro and microaverage results of the proposed

frameworks and the baseline methods in Table 2

From Table 2 we first observe that on the long

tailed dataset the methods finetuned on teacher

generated rationales CoT can significantly out

perform the groundtruth finetuning method Ran

dom Finetune which emphasizes the necessity of

teachergenerated reasoning steps in the KD

Among all sequencelevel KD methods our pro

posed BalDistill N and BalDistill A achieve the

best average performance across various datasets

on macroaverages which obtain an average rel

ative improvement of 224 and 681 respec

tively compared to the Random Finetune CoT base

line The performance boost in BalDistill N im

plies the effectiveness of replacing the naive bal

ancing policy with adaptive balancing

Moreover we note that the Duplicate Finetune

CoT baseline fails to compete with the Random

Finetune CoT method in most cases which indi

343 289 224 108 35

Domain Proportion 045048050053055058F1a AbstractiveQA

294 09 05 03 02 01

Domain Proportion 000020040060080100F1 b Reuters

1861751551201168854483028

Domain Proportion 060070080090Acc

c MultichoiceQA

237 181 174 109 108 96 95

Domain Proportion 010020030AccRandom

Duplicate

BalDistill A d Math

Figure 2 Performance of proposed method and base

lines on different domains Xaxis represents the pro

portion of each domain ranked from head to tail do

mains Our proposed BalDistill method can achieve

comparable results on head domains and outperform the

baseline method on the tail domains

cates that simply duplicating the input from the tail

domains to ensure balanced data cannot address

the underlying imbalanced data complexity

To perform a detailed analysis of our framework

we visualize the F1 or accuracy score for each do

main of the BalDistill N method and two baseline

methods Random Finetune CoT and Duplicate

Finetune CoT in Figure 2 with the xaxis repre

senting the proportion of each domain in the dataset

in descending order From Figure 2 our proposed

method can achieve comparable results in the head

domains left side of the figure but substantially

outperform the baseline methods in the tail domains

right side of the figure This observation verifies

our expectation in Section 3 where the balancing

policy increases performance in the tail domain

and the active learning part improves the data ef