KBS: A Korean Benchmark Suite for Foundation Model Evaluation
A Korean benchmark suite spanning six capabilities, with task-specific adaptation and analysis of how evaluation settings shape model scores.
Profile photo
Image unavailable
Yeonkyoung So
소연경
Ph.D. Student in Data Science
Graduate School of Data Science
Seoul National University
My research centers on benchmarking and evaluating language models. I aim to build evaluations that reveal model capabilities and limitations beyond aggregate scores.
Research interests
- Benchmark Design
- Model Evaluation
- Language Understanding
I study how benchmark design, evaluation protocols, and scoring methods shape our understanding of language models. My work includes Korean language benchmarks, sentence-level negation, and answer-choice sensitivity.
More broadly, I am interested in the reliable and responsible use of AI in real-world settings.
Selected publications
Google Scholar2026
Equal contribution
Thunder-DeID+: Boundary-Aware Korean De-identification from Court Judgments to Broader Sensitive Documents
Boundary-aware Korean de-identification with a refined PII taxonomy and tokenizer-independent evaluation, extending from court judgments to broader sensitive documents.
Thunder-KoNUBench: A Corpus-Aligned Benchmark for Korean Negation Understanding
A corpus-aligned benchmark for Korean negation understanding, with analyses of model behavior and the effects of fine-tuning.
Choices Speak Louder than Questions
An analysis of choice sensitivity in multiple-choice evaluation and a scoring method that isolates the contribution of the question.
Thunder-NUBench: A Benchmark for LLMs’ Sentence-Level Negation Understanding
A sentence-level negation benchmark that distinguishes standard negation from local negation, contradiction, and paraphrase.
Experience
–
CHAMP
Research Assistant · Thunder Research Group
Worked on LLM evaluation methods and dataset development.
–
NCSOFT
AI Learning Data Construction Assistant
Assisted in the development and structuring of AI training data.
–
ENlighten
Data Team Intern
Conducted data analysis and module development.
Education
– Present
Ph.D. in Data Science
In progressSeoul National University
–
Master’s in Data Science
Seoul National University
ThesisNeVer: A Comprehensive Dataset for Negation Handling in Language Models
Developed a human-reviewed dataset to evaluate and improve negation understanding in language models.
–
Bachelor’s in Western History
Seoul National University
Combined minor: Data Science for the Humanities
Undergraduate thesis
Secularization in the United States and the Reaction of Christian Fundamentalism: Focusing on Billy Sunday in the Early 20th Century
Western History · Written in Korean
A historical study of Billy Sunday’s fundamentalist movement and its responses to secularization in the early twentieth-century United States.
Selected awards
최우수상
4기 구름톤
Top Excellence Award
goormthon · 4th edition
우수상
2023 관광데이터 활용 공모전
Excellence Award
2023 Tourism Data Utilization Contest
Certifications
- 빅데이터분석기사
- SQLDSQL 개발자
- ADsP데이터분석 준전문가
Words & music
伏久者飛必高
오래 엎드려 있던 새는 반드시 높이 난다.
More quotes
When the going gets tough, the tough get going.
BGM
I Want to Break Free Queen
A song and a state of mind.
A few things about me
I enjoy traveling (38 countries so far) and have played 130+ escape room games.
I like learning languages. I’ve studied Chinese, Japanese, and French (DELF B1), though English is the one I keep working on.
I’ve taken piano, violin, and flute lessons, and I enjoy singing and watching films.
I’m not particularly sporty, but I once represented my school in a table tennis tournament. I’m not especially outdoorsy either, but I do like hiking.
There’s always something else I’d like to try. (the list keeps growing )