2024 자동차 데이터 분석 경진대회 : 생성 AI 프롬프트 엔지니어링

뭐가 틀린거죠?

2024.10.02 13:55 1,598 조회

system,user

"# my prompt ..

....


# aaaa ...


# bbb ..

"","

ID TEST_00

title Beach Profile Data Collected from Madeira Beach, Florida (January 15, 2021)

notes This dataset, prepared by the U.S. Geological Survey (USGS) St. Petersburg Coastal and Marine Science Center (SPCMSC), provides beach profile data collected at Madeira Beach, Florida. Data were collected on foot by a person equipped with a Global Positioning System (GPS) antenna affixed to a backpack outfitted for surveying location and elevation data (XYZ) along pre-determined transects. The horizontal position data are given in the Universal Transverse Mercator (UTM) projected coordinate system, Zone 17 North (17N), referenced to the North American Datum of 1983 (NAD 83); the elevation data are referenced to the North American Vertical Datum of 1988 (NAVD 88), GEOID12B.


ID TEST_01

title Nota media de la nota de admisi?n (cohorte de nuevo ingreso en el SUE) por forma de admisi?n, ?mbito de estudio y sexo (universidades p?blicas presenciales)

notes Tabla de EDUCAbase Nota media de la nota de admisi?n (cohorte de nuevo ingreso en el SUE) por forma de admisi?n, ?mbito de estudio y sexo (universidades p?blicas presenciales). Indicadores de los NO beneficiarios de becas generales de la AGE y Pa?s Vasco en estudios de Grado. Resultados por universidad. Anual.

...

"




이와 같이 제출 했습니다...

근데 다 0점이 나오네요.

로그인이 필요합니다
0 / 1000
NN_is_all_you_need
2024.10.02 15:00

출력이 아마 40개행 마다 0과 1로 출력되지 않아서 그럴것 같습니다

꿀시럽
2024.10.02 16:29

디버깅 불가한 프로그래밍 같은 건가요... 
연습할 수 있는 곳이 있는 것도 아니고 충분히 실패 reason을 아웃풋 할 수 있을법도 한데 아쉽네요

IIllIIllIIll
2024.10.02 16:47

챗지피티 가서 하거나 직접 API로 돌려보면 연습할수있을듯여

에스삐
2024.10.03 02:15

제출하기 전에 0과 1이 40개가 출력되는지 확인하고 제출해보세요

월롱이
2024.10.03 02:16

저는 0 과 1 로만 40 행 정확히 출력하는데도 제출하면 0점이네요..

BG01882
2024.10.03 14:57

같은 프롬프트로, 항상 40이 나오지 않았습니다.
프롬프트를 바꿔보시어 40이 더 잘 나오도록 유도하시는게 좋을 것 같습니다.

weall
2024.10.04 01:16

행마다 출력이면 
1
0
1
이런식 아님? 
뭔 제출 할때마다 다르게 나오는건가

월롱이
2024.10.04 14:34

저희가 모델에 직접 접근할 수 없는 특성상 시드를 고정할 수 없고, 입력마다 다르며, 39, 41개 행으로 출력되는 경우가 허다하네요. 정확한 행의 답변이 나오도록 요구하면서도 성능을 올리는 프롬프트가 관건인것 같습니다. 
간겅성이 좋은 프롬프트를 사용하고, api 를 통해 확인하더라도 제출했을때는 결국 0점이 나오는 경우가 많습니다. 프롬프트 뿐 아니라 운도 함께 작용할 수 있다고 할 수 있죠

byc3230
2024.10.09 19:57

강건성을 확보한 api 테스트 반복 테스트 까지 진행했음에도 불구하고 그럴수도 있을까요? 쉽지 않네요 ㅠㅠ

BG01882
2024.10.09 22:45

gpt-3.5-turbo에 temp 0.4로 하셨던걸까요? 프롬프트마다 강건성이 달라지긴 하는데 적어도 리더보드 제출 시 5번 중 3번 이상은 유효점수가 나왔습니다.

minu_13
2024.10.10 11:02

0과 1이 40개가 출력되는지 확인하고 제출하셔야 할 것 같습니다 !