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 롤링 시청 페르소나 (Task 1)

LGUplus 컨텍스트 이해 POC용 소형(350M) 모델 학습 데이터셋. 교사 모델 Qwen3-235B-A22B-Instruct-2507-FP8로 생성.

롤링 페르소나 갱신: 어제까지의 페르소나 + 오늘 하루(1일분) 시청이력 → 오늘을 반영한 갱신 페르소나. 주간 스텝·코드 계산 없이 매일 순수 일일 롤링(컨텍스트는 페르소나가 들고 다님).

포맷 (chat messages)

각 row = 하나의 (persona, 날짜) 갱신 스텝.

컬럼 설명
persona_id 시청자 ID
date / dow 날짜 / 요일
mode coldstart(첫날) / daily
messages [system, user, assistant] (content=문자열)
  • system: 태스크 지시(페르소나 갱신 분석가)
  • user: [이전 페르소나] + [오늘 시청 기록](시청시간·채널·제목·EPG장르·EPG설명·자막요약)
  • assistant(타깃): 갱신된 페르소나 = [장기 성향](프로즈) + [최근 성향](프로즈) + [시청 프로그램]

페르소나 구조

  • 장기 성향: 누적 시청 패턴(장르·시간대·주제·정서) 기반 지속 취향.
  • 최근 성향: 오늘·최근 며칠 관심 + 새로 보기 시작(신규)·뜸해진(감소) 프로그램을 이름과 함께 프로즈로.
  • 시청 프로그램 표: 자리잡은 고정 루틴(상승·안정)만. 신규·일회성·뜸해진 건 표에 넣지 않고 최근 성향에 서술 → 표가 작게 유지됨(median 3).

스플릿

  • train: 9,900 persona (199,238 step)
  • test: 100 persona (1,997 step) — 고정 시드(42) 무작위, persona 단위 분리(누수 없음)

학습 (assistant-only loss)

messages가 문자열 content 표준형이라 apply_chat_template로 바로 사용. TRL SFTConfig(assistant_only_loss=True) + {% generation %} 포함 ChatML 템플릿으로 assistant 구간만 loss.

출처

  • 시청이력: HBKenerzai/LGUplus_viewing_history
  • EPG 요약/장르: HBKenerzai/LGUplus_epg_summary
Downloads last month
20