kor-persona-survey-sample

A deliberately reduced demonstration sample (800 triplets) of the synthetic training data behind dobstudio/kor-persona-survey-7b: Korean persona profiles answering survey / balance-game questions in character, labeled by a Qwen2.5-72B-Instruct teacher.

Shape: 16 questions × 50 distinct personas each. The sample is intentionally question-narrow and persona-wide so that its most visible property is the model's core claim — the same question, answered differently by different personas.

Fields

field	description
persona_uuid	ID in nvidia/Nemotron-Personas-Korea
persona_profile	The exact profile text injected into the prompt (demographics + narrative)
question	Survey question (Korean, binary choice)
options	The two answer options
category	One of 8 non-sensitive topics (food, lifestyle, consumption, travel, leisure, work, values, digital)
response	Teacher label: {choice, confidence, reason}

How it was made

(persona sampled from Nemotron-Personas-Korea) + (question from a curated private question bank) → Qwen2.5-72B-Instruct-AWQ generates the in-character response → quality filtering (format validity, option match, confidence range, reason de-duplication; 95.2% pass rate).

Not released: the full 28k-triplet training set, the 515-question bank, evaluation (held-out) questions and their labels, and the population-aggregation pipeline. Politically/socially sensitive topics were excluded by design.

한국어 요약

dobstudio/kor-persona-survey-7b 학습 데이터의 의도적으로 축소된 시연용 샘플(800건)입니다. 한국인 가상 인물 프로필이 설문/밸런스게임 문항에 인물의 입장에서 답한 (페르소나, 문항, 응답) 데이터로, 응답은 Qwen2.5-72B 교사 모델이 생성했습니다.

• 구성: 문항 16개 × 문항당 서로 다른 인물 50명 — "같은 질문에 인물마다 다른 답" 이라는 모델의 핵심 성질이 그대로 보이도록 문항은 좁게, 인물은 넓게 설계했습니다.
• 비공개: 전체 학습 데이터(2.8만 건), 문항 뱅크 515개, 평가(held-out) 문항과 라벨, 모집단 집계 파이프라인. 정치 등 민감 주제는 애초에 배제했습니다.
• 인물은 합성이며, 응답은 실제 사람의 의견이 아니라 LLM의 판단입니다.

Attribution & license

• This dataset: CC BY 4.0.
• Persona profiles derive from nvidia/Nemotron-Personas-Korea (CC BY 4.0, © NVIDIA).
• Responses were generated with Qwen2.5-72B-Instruct-AWQ (Qwen LICENSE, Alibaba Cloud) — Built with Qwen.
• Personas are synthetic; responses reflect LLM judgments, not real people's opinions.