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_persona_4task_sft_v2

LG U+ 페르소나 350M(ternary QAT) 모델의 rolling 기반 SFT 학습 데이터 (v1의 rolling 버전). 각 행 = {task, prompt, completion}. 학습 시 loss는 completion(<|gen|> 뒤 + <|im_end|>)에만.

v1(LGUplus_persona_4task_sft)과의 차이:

  • 일일(day)+주간(week) 페르소나 태스크를 rolling 페르소나 갱신 태스크 하나(persona)로 통합 → 4태스크 → 3태스크.
  • 추천(chan)·키워드(kw)의 페르소나 컨텍스트를 주간+직전일 대신 전일까지 갱신된 rolling 페르소나 텍스트로 대체.

3태스크

task 토큰 입력(prompt) 출력(completion)
persona <|persona|> [이전 페르소나] + [오늘 시청 기록] 갱신된 페르소나 텍스트([장기 성향]+[최근 성향]+[시청 프로그램] 표)
chan <|chan|> rolling 페르소나 + 최근2h 시청 + 후보채널(20개, 유니버스 distractor로 증강·셔플) {"channels":[≤5]}
kw <|kw|> rolling 페르소나 + 후보키워드(≥2) [최근시청 미포함] {"keyword":"top1"}

구성 (서브샘플: persona 전량, chan/kw 각 100,000)

  • data/train.jsonl (399,238) : persona 199,238 + chan 100,000 + kw 100,000
  • data/val.jsonl (5,997) : persona 1,997 + chan 2,000 + kw 2,000

chan은 후보채널≥2, kw는 후보키워드≥2(실제 리랭크)만 포함. 라벨은 항상 후보의 부분집합(검증됨).

스키마

{"task": "chan",
 "prompt": "<|startoftext|><|chan|>[페르소나]\n[장기 성향]...\n[최근 성향]...\n[시청 프로그램]...\n최근:...\n후보채널:KBS2,...<|gen|>",
 "completion": "{\"channels\":[\"KBS2\",...]}<|im_end|>"}

원천 데이터셋 (rolling)

주차: wk0=5/316/6, wk1=6/713, wk2=6/14~20 (연속 3주). chan/kw의 페르소나 컨텍스트는 각 행 persona_date(전일까지 갱신) 스냅샷.

Downloads last month
24