You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

LGUplus 키워드 선택 (Task 2)

LGUplus 컨텍스트 이해 POC용 소형(350M) 모델 학습 데이터셋. 교사 Qwen3-235B-A22B-Instruct-2507-FP8로 생성. Task 1: 롤링 페르소나 와 같은 시청자·동일 test 분리.

키워드 선택: 페르소나 + 최근 2시간 시청 + 후보 키워드 → 지금 추천하기 가장 좋은 유튜브 검색어 1개.

포맷 (chat messages)

각 row = 하나의 (persona, 날짜, 2시간 블록).

컬럼 설명
persona_id / date / block 시청자 / 날짜 / 2시간 블록 시작시각
n_watched 그 블록에 실제 시청한(정답 후보) 키워드 수(2~4)
pick_in_watched 교사가 고른 키워드가 시청 키워드였는지(distractor 아님)
messages [system, user, assistant] (content=문자열)
  • system: 키워드 큐레이터 지시
  • user: [페르소나](Task1 산출) + [최근 2시간 시청] + [후보 키워드](번호목록)
  • assistant(타깃): 선택된 키워드 1개

후보 구성

  • 후보 8개 = 그 블록에 15분+ 시청한 프로그램들의 키워드(정답 후보, 2~4개) + distractor(같은 시청자의 다른 시간대 시청 키워드 → 부족분은 카탈로그) 패딩.
  • 프로그램별 키워드는 각 프로그램 요약에서 생성(유튜브 검색어 스타일, 3~6단어). 스타일 참고: jungsanghyun/subtitle-summary-{lgu,dataset}.
  • 교사가 시청 후보를 고른 비율 ≈ 92.5%(나머지는 페르소나에 더 맞는 distractor 선택).

스플릿

  • train: 177,708 (9,900 persona) · test: 1,809 (100 persona) — persona 단위 분리(Task1과 동일 test 셋, 누수 없음)

출처

  • 시청이력/페르소나: HBKenerzai/LGUplus_viewing_history, Task1 페르소나
  • EPG 요약: HBKenerzai/LGUplus_epg_summary
Downloads last month
7