README.md

metadata

license: apache-2.0
language:
  - ko
base_model: Qwen/Qwen2.5-7B-Instruct
datasets:
  - nvidia/Nemotron-Personas-Korea
pipeline_tag: text-generation
tags:
  - korean
  - persona
  - role-play
  - survey-simulation
  - distillation
  - qlora

kor-persona-survey-7b

[한국어 요약이 아래에 있습니다 / Korean summary below]

A 7B model distilled to answer survey and preference questions as a specific Korean persona, given a structured persona profile. Fine-tuned from Qwen2.5-7B-Instruct via QLoRA on synthetic (persona, question, response) triplets labeled by a Qwen2.5-72B-Instruct teacher, using persona profiles from nvidia/Nemotron-Personas-Korea.

Research question: How much of a 72B teacher's persona-conditioned survey-response capability can be transferred to a 7B student through synthetic-triplet distillation alone?

What it does

Input: a Korean persona profile (demographics + narrative fields) and a survey/balance-game question. Output: a structured JSON response in character:

{"choice": "짬뽕", "confidence": 0.8, "reason": "얼큰한 국물 없이는 식사가 허전해서"}

Intended for: persona-conditioned response simulation research, synthetic survey data generation, Korean role-play agent studies.

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "dobstudio/kor-persona-survey-7b"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

persona = """- 성별/나이: 여성, 58세
- 지역: 부산광역시
- 직업: 식당 운영
- 음식 성향: 매운 음식을 즐기며 직접 담근 김치에 자부심이 있음"""

question = "짜장면 vs 짬뽕, 하나만 고른다면?"

messages = [
    {"role": "system", "content": f"당신은 아래 인물입니다. 이 인물의 입장에서 설문에 답하세요.\n{persona}\n\n반드시 JSON으로만 답하세요: {{\"choice\": ..., \"confidence\": 0.0~1.0, \"reason\": \"한 줄 이유\"}}"},
    {"role": "user", "content": question},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=128)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

With vLLM (recommended for batch simulation):

from vllm import LLM, SamplingParams

llm = LLM(model="dobstudio/kor-persona-survey-7b", max_model_len=2048)
outputs = llm.chat([messages], SamplingParams(temperature=0.7, max_tokens=128))
print(outputs[0].outputs[0].text)

Training

Base model	Qwen/Qwen2.5-7B-Instruct (Apache-2.0)
Method	QLoRA SFT (4-bit NF4, LoRA r=16 / α=32, all linear layers, loss on response only), 2 epochs, effective batch 48, cosine lr 1e-4
Teacher	Qwen/Qwen2.5-72B-Instruct-AWQ — generated response labels; disclosed for transparency
Training data	27,999 synthetic triplets (28,571 after quality filtering, 95.2% pass rate from 30,000 raw): (persona from Nemotron-Personas-Korea, question from a curated 515-question bank, teacher-labeled response). Question topics: food, lifestyle, consumption, travel, leisure, work, values, digital — politically/socially sensitive topics excluded
Data release	An 800-triplet sample (16 questions × 50 distinct personas each): https://huggingface.co/datasets/dobstudio/kor-persona-survey-sample. The full triplet set, question bank, and the population-aggregation pipeline are not released
Hardware	RTX 3090 (24GB) ×4 workstation; training ran as 3-GPU DDP, 2.5 h

Evaluation

All Model-level metrics are measured on held-out questions never seen in training, and are reproducible with the released model and the prompt format above.

Protocol: 51 held-out questions × 20 evaluation personas (disjoint from training personas), temperature 0.7 with fixed per-request seeds.

Model-level — standalone model metrics (reproducible)

Metric	Base Qwen2.5-7B	This model	Teacher 72B
Format compliance (valid JSON + exact option match)	95.5%	98.5%	97.0%
Teacher agreement (held-out)	72.3%	79.4%	96.1%¹
Response consistency (5-run majority reproducibility)	98.7%	97.6%	—
Persona sensitivity (TV distance vs. no-persona baseline)	0.356	0.356	—
Position bias (first-option preference)²	—	+1.7%	—
Age-conditioning sensitivity (contrastive pairs)³	—	+20.8%p	—

¹ Teacher self-agreement across two independent samplings — the effective ceiling for teacher agreement under temperature 0.7. The distilled model reaches 82.6% of that ceiling (79.4 / 96.1), up from 75.2% before distillation. ² Estimated as (first-option share, normal order + first-option share, flipped order − 1) / 2 over held-out questions; 0 = unbiased. Choice consistency under option-order flip is 90.7%, statistically at the sampling-noise floor (91.4%). ³ Flipping only the age field (27↔67) in demographics-only profiles changes the chosen option on 29.4% of questions vs. an 8.6% same-profile resampling noise floor — net +20.8%p of genuine age conditioning.

Scaling ablation

We verified that the released configuration saturates this method: doubling the training data (56k triplets from 6,000 personas) left held-out teacher agreement unchanged (79.3%) despite improving validation loss (0.391 → 0.357), and a third epoch moved it only within sampling error (80.1%, n≈980). Closing the remaining gap to the teacher self-agreement ceiling likely requires methodological changes (multi-sample distribution distillation, preference optimization) rather than more of the same data.

System-level reference — not yet measured

The authors plan to report population-level accuracy (MAE of predicted answer distributions vs. real Korean survey results, using stratified sampling + post-stratification against KOSIS census margins) in a future update. Those numbers will depend on a private aggregation pipeline and will be reported for context only.

Limitations and ethical considerations

• Synthetic personas are not real people. Outputs simulate what a fictional profile might answer based on LLM priors — they are not measurements of actual Korean public opinion and must not be presented as such.
• Joint-distribution distortion in the source data. Nemotron-Personas-Korea matches real Korean marginal distributions (age, sex, occupation) but its joint distributions (e.g., age×occupation×region) deviate from reality (see arXiv:2606.12433). Segment-level conclusions drawn from raw persona samples are unreliable without statistical correction.
• Stereotype risk. The teacher may role-play personas stereotypically, and distillation inherits this. Persona-sensitivity metrics partially quantify conditioning, not fairness.
• Domain bound. Trained on non-sensitive preference/lifestyle questions. Behavior on political, medical, or otherwise sensitive questions is untested and out of scope.
• Outputs are in Korean; other languages are untested.

Attribution & license

• Model weights: Apache-2.0.
• Persona profiles: nvidia/Nemotron-Personas-Korea, CC BY 4.0 — © NVIDIA, used with attribution as required.
• Base and teacher models: Qwen2.5 family (Apache-2.0), Alibaba Cloud.

---

한국어 요약

한국인 가상 인물 프로필을 주면 그 인물의 입장에서 설문/밸런스게임 문항에 구조화된 JSON으로 답하는 7B 모델입니다. Qwen2.5-72B 교사 모델이 생성한 (페르소나, 문항, 응답) 합성 데이터로 Qwen2.5-7B를 QLoRA 증류했습니다.

• 연구 질문: 72B의 페르소나 조건부 응답 능력이 증류만으로 7B에 얼마나 이식되는가
• 평가: 학습에 쓰지 않은 문항(held-out)에서 포맷 준수율, 교사 일치율, 응답 일관성, 페르소나 감수성을 측정 — 공개된 모델만으로 재현 가능
• 한계: 가상 인물의 응답은 실제 여론이 아닌 추정이며, 원본 데이터의 결합분포 왜곡으로 세부 집단 분석에는 통계 보정이 필요합니다. 민감 주제는 학습에서 제외했습니다.
• 모집단 집계 파이프라인(층화 표본추출·사후층화 가중)은 비공개이며, 본 공개물은 개인 페르소나 역할극 기능만 제공합니다.

---