Obfuscated Korean Review Restoration and Generative AI Competition

Algorithm | Montly Dacon | NLP | Generate AI | LLM | F1 Score

  • moneyIcon Prize : DASCHOOL Pro Subscription
  • 2025.01.06 ~ 2025.02.28 09:59 + Google Calendar
  • 727 Users Completed

 

private 4th 0.97478) Phoneme based Gemma2 2b Encoder

2025.02.28 10:56 1,084 Views language

다들 수고하셨습니다.

저는 대회 매트릭상 고정된 위치의 출력을 생성할 수 있는 인코더 방식이 적합하다고 생각하고, 단어나 글자단위보다는 자음/모음 단위의 교정이 필요하다고 생각했습니다.
gemma모델이 한글과 한글을 초중종성으로 나눴을때의 이해도가 높다고 판단되어서 gemma 모델을 encoder로 변환하여 대회를 진행했습니다.
public 기준 0.972 -> 0.974(+데이터증강) ->0.975(+후처리) 순으로 성능이 증가했습니다.

PDF
Code
Login Required
0 / 1000
whybe
2025.03.03 23:52

LLM2Vec을 보고 이런 생각을 하셨다는 게 정말 신기하네요,, 좋은 아이디어 공유해주셔서 감사합니다!

파이썬초보만
2025.03.04 08:22

잘 읽어주셔서 감사합니다!

힐링이필요해
2025.03.05 09:17

아이디어가 진짜 좋네요. 덕분에 정말 많이 배웠습니다! 
감사합니다~

파이썬초보만
2025.03.06 10:18

감사합니다 ㅎㅎ