Earticle

초록

영어

Small Language Models are competitive without large-scale infrastructure, their performance is highly contingent on prompt design. This study analyzes the sensitivity of BitNet b1.58-2B-4T to label exposure and fewshot exemplar composition on a 36-class medical query classification task. We generated 504 items consisting of 6 direct and 8 indirect questions for each disease and after removing cross-exemplar leakage the final evaluation set contained 494 items. With no parameter updates, 0/1/2/5/10- shot prompting was evaluated using Accuracy. Under the nolabel- exposure setting accuracy increased as more exemplars were provided. However, these gains were accompanied by growing prediction concentration on exemplar labels. In contrast with label-exposure, zero-shot achieved the highest accuracy, while the inclusion of exemplars reduced accuracy and amplified label bias. These results show that the structure of the prompt tends to shift few-shot effects from beneficial to detrimental. This highlights the importance of controlled prompt design and domain-adaptive training to ensure trustworthy performance.

목차

Abstract
I. INTRODUCTION
II. METHODOLOGY
III. EXPERIMENTS AND RESULTS
A. Experiment Settings
B. Experiment Results
IV. DISCUSSION AND CONCLUSION
ACKNOWLEDGMENT
REFERENCES

저자

• Sihyung Kim [ Department of Computer Engineering The Catholic University of Korea Bucheon, South Korea ]
• Jaehyun Cha [ Department of Computer Engineering The Catholic University of Korea Bucheon, South Korea ]
• Siyoung Kim [ Department of Computer Engineering The Catholic University of Korea Bucheon, South Korea ]
• Yoojoong Kim [ School of Computer Science and Information Engineering The Catholic University of Korea Bucheon, South Korea ] Corresponding Author

참고문헌