text
stringlengths 0
89.3k
|
|---|
mentation could achieve the best average results
|
even better than the naive balancing method We
|
conjecture that since we do not verify the correct
|
ness of teachergenerated samples and rationales
|
in tail domains While teachergenerated samples
|
induce more knowledge more synthetic data can
|
lead to more inevitable noise Adaptive balancing
|
achieves the best tradeoff between inducing more
|
knowledge and less noise in the tail domains
|
53 Generalization Analysis
|
The ablation study demonstrates the effectiveness
|
of the active learning and adaptive balancing Then
|
we ask whether our proposed method is robust
|
enough to experiment with different hyperparame
|
ters student models or budget settings
|
531 Generalizations on Student models
|
Method R52 Reuters AbsQA MCQA Math
|
Llama3 Budget Setting 1
|
Random FT CoT 5970 2735 5257 7609 1520
|
Active FT CoT 5949 2975 5314 7664 1561
|
BalDistill A 5893 3295 5320 7717 1742
|
Llama2 Budget Setting 1
|
Random FT CoT 4983 2397 4626 5869 343
|
Active FT CoT 4688 2406 4707 5868 382
|
BalDistill A 5833 2551 4755 5914 421
|
Llama3 Budget Setting 2
|
Random FT CoT 6488 3342 5371 7292 1519
|
Active FT CoT 6454 3133 5305 7626 1591
|
BalDistill A 6584 3277 5349 7711 1759
|
Llama2 Budget Setting 2
|
Random FT CoT 5645 2375 4895 5891 384
|
Active FT CoT 5316 2712 4827 5920 352
|
BalDistill A 5817 2707 4945 5864 454
|
Table 4 Effects of student model scales and budget
|
numbers Macroaverages the proposed and baseline
|
method results when considering Llama2 and Llama3
|
as student models with varying two budget settings
|
We first evaluate whether our method could be
|
generalized to student models with different rea
|
soning abilities or with different budget numbers
|
In this part we additionally evaluate our BalDistill
|
A on Llama27B models which have a smaller
|
model size and fewer tokens in two budget set
|
tings the details of each dataset are in Table 1
|
We present the finetuning results of our proposed
|
framework and baseline methods on the Llama2
|
and Llama3 student models in Table 4
|
BalDistill exhibits robust improvement with var
|
4 6 8
|
Stage number048050052054056058F1
|
BalDistill A
|
Randoma AbstractiveQA
|
4 6 8
|
Stage number025028030033035038F1
|
BalDistill A
|
Random b Reuters
|
Figure 4 Influence of stage number choices on BalD
|
istill across two datasets Our proposed method consis
|
tently obtains better results than the random finetune
|
baseline method with varying stage numbers
|
ious budget settings or student models From
|
Table 4 we observe that finetuning with the
|
Llama38B student model leads to much better per
|
formance than the Llama27B model especially on
|
tasks with complex reasoning Math Multichoice
|
QA indicating that the student with a larger model
|
size or a better reasoning ability will yield better
|
finetuning results This observation is consistent
|
with previous findings in Ho et al 2022 Hsieh
|
et al 2023 Our BalDistill A consistently outper
|
forms other baseline methods on both Llama27B
|
and Llama38B as student models in most cases
|
under two budget settings which also verifies the
|
generalizability of our BalDistill A on different
|
student models or different budget numbers
|
532 Sensitivity Analysis
|
We next investigate how the choice of stage number
|
Kwill influence the performance of our frame
|
work We experiment with the same setup as in
|
Section 51 but with varying stage numbers among
|
358 We visualize the results macroaverages
|
of BalDistill A and the baseline method Random
|
Finetune CoT in Figure 4 and 6 in Appendix
|
Figure 4 shows that the finetuning results of
|
BalDistill A could be affected by the stage num
|
ber to some extent but our proposed method can
|
consistently outperform the baseline method with
|
different stage numbers demonstrating the effec
|
tiveness and robustness of BalDistill A
|
6 Conclusions
|
In this paper we propose a novel framework BalD
|
istill to enhance performance on longtail datasets
|
in the current teacherstudent knowledge distil
|
lation process Our framework is a multistage
|
pipeline and at each stage we call the student mod
|
els to actively select the representative examples
|
from head domains while prompting the teacher togenerate synthetic examples for tail domains With
|
a fixed budget restriction for calling the teacher
|
our extensive empirical evaluations show that our
|
framework can significantly increase finetuning re
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.