Fine-tuning DataText

Multilingual General-Domain QA Text Dataset

Type
QA Dataset
Domain
Humanities and Social
Language
KoreanHindiIndonesianArabicThaiBengaliArabic (Egypt)Japanese

Overview

A parallel multilingual question-answering dataset covering economics, politics, and society, reviewed and corrected by professional annotators. Answers open by framing what the question is really asking, then work through the substance in ordered points, and where the topic has a trade-off rather than a settled answer they set out both sides — the gains and the costs of stronger IP protection, the tension between non-intervention and humanitarian intervention — instead of presenting one position as consensus.

Much of the material is anchored in the Korean context: how credit scores are set in Korea, why the model taxpayer award exists, how conspicuous consumption plays out in Korean society. The same question and answer are then carried across eight languages, so the dataset transfers Korean social and economic context into each language rather than replacing it with a local equivalent. Questions are short relative to their answers and arrive in several registers, from a casual question between friends to an essay prompt, so the set trains both explanatory and analytical responses.

Applicable Areas

  • Supervised fine-tuning of general-purpose assistant models
  • Training balanced, two-sided responses on topics that involve trade-offs
  • Cross-lingual evaluation of answer framing and factual grounding
  • Localizing Korean social, economic, and policy context into other languages

Dataset Preview

indexcategoryq_idquestionanswerquestion_token_countanswer_token_count
1437EconomicsECO_01437친구 부모님이 모범납세자상을 받으셨다던데, 납세는 당연한 의무인데 왜 상을 주는 거야? 그리고 이 상을 받으면 어떤 혜택이 있어?모범납세자상은 단순히 납세 의무를 이행했다고 주는 것이 아니라, 성실하고 모범적인 납세 행위를 인정하고 격려하기 위한 제도입니다. (…)49629
283PoliticsPOL_00283지적 재산권 보호 강화가 한국의 기술 수출에 미치는 긍정적·부정적 영향을 분석하고, 이를 최적화하기 위한 무역 전략을 간단히 제시하시오.대한민국의 지적 재산권(IP) 보호 강화는 기술 수출에 긍정적인 영향을 미치는 여러 가지 방면이 있습니다. IP 보호는 창업과 혁신을 장려하고, 기술 소유자에게 안정적인 수익원을 제공합니다. (…)46724
1285SocialSOC_01285베블런(Veblen) 효과로 설명되는 한국의 과시적 소비 문화가 사회 전반의 경제적 안정성에 미치는 부정적 영향은 무엇인가?베블런 효과는 미국의 경제학자 토스테인 베블런이 제시한 개념으로, 가격이 오를수록 수요가 증가하는 현상을 설명합니다. 한국 사회에서 이러한 과시적 소비 문화가 경제적 안정성에 미치는 부정적 영향은 다음과 같습니다: (…)38568
  • Please review the data specifications and actual samples in advance.
  • The complete data shown in this preview is available in the sample download.

How to create

Collection

Questions are sourced across economics, politics, and society and assigned a category-prefixed identifier (ECO_ / POL_ / SOC_) so that domain balance is auditable at a glance. Registers are collected deliberately — the casual phrasing of someone asking a friend, the neutral phrasing of a general inquiry, and the directive phrasing of an essay prompt — because the same topic draws a different answer shape depending on how it is asked.

Answers are drafted to a consistent shape: open by establishing what the question turns on, then work through the substance in ordered points, and where the topic involves a trade-off, set out both the upside and the downside rather than resolving it. Korean institutional and social context is stated explicitly, including the institution or agency involved, so that an answer remains interpretable once translated. Each answer is then localized into the eight target languages, with Korean context, institution names, and concept terms carried across intact.

Validation

Professional annotators verify factual claims against primary sources, with particular attention to figures, institution names, and the mechanics of Korean systems, since these are the parts most likely to be stated approximately. Reviewers reject answers that embed a normative judgment as though it were fact, and answers that present one side of a trade-off while the question asked for both.

Localized versions are then checked back against the Korean source so that no language version drops a counterargument, flattens a hedge into a claim, or substitutes a local institution for the Korean one named in the source. Token counts are recorded per locale after review, which surfaces cases where one language version is materially shorter than its source and therefore likely to have lost a point.

Flitto Curation Data Flitto Curation Data