Fine-tuning DataText

Multilingual STEM QA Text Dataset

Type
QA Dataset
Domain
IT and TechScience and Engineering
Language
KoreanHindiIndonesianArabicThaiBengaliArabic (Egypt)Japanese

Overview

A parallel multilingual question-answering dataset spanning science, technology, engineering, and mathematics, reviewed and corrected by annotators with a STEM major. Answers are worked solutions rather than verdicts: they restate the problem, name the governing equation or principle, substitute the given values, and carry the derivation through to a result, with each stage of the working left visible so the answer can be verified rather than trusted.

The same question and answer are carried across eight languages. Equations are written in LaTeX and preserved verbatim in every language version, and English technical terms are kept as they are actually used by practitioners instead of being forced into a native equivalent — a Korean question about a causal system asks for its unit impulse response in those words. Questions arrive in two registers, from the formal phrasing of a problem set to the conversational phrasing of a working engineer asking a colleague, so the set trains both instructional and consultative responses.

Applicable Areas

  • Supervised fine-tuning of LLMs for technical and scientific reasoning
  • Evaluating derivation validity, numeric accuracy, and unit handling in generated answers
  • Cross-lingual consistency testing of STEM terminology and notation
  • Building STEM tutoring and engineering assistant applications

Dataset Preview

indexcategoryq_idquestionanswerquestion_token_countanswer_token_count
103ScienceSCI_00103플루오린화리튬(LiF)과 플루오린화세슘(CsF)의 격자 에너지를 계산하고 비교하세요. Madelung 상수는 LiF: 1.750, CsF: 1.763입니다. 리튬 이온의 반지름은 0.076 nm, 세슘 이온의 반지름은 0.167 nm입니다. 어떤 화합물이 더 강한 격자...플루오린화리튬(LiF)과 플루오린화세슘(CsF)의 격자 에너지를 계산하고 비교해 보겠습니다. (…) 격자 에너지를 계산하기 위해 Born-Landé 방정식을 사용할 수 있습니다: $$ U = -\frac{N_AMe^2}{4\pi\epsilon_0r_0} (1 - \fr...113667
571EngineeringENG_00571input이 x(n)이고 output이 y(n)인 causal system 이 있을 때, 다음과 같은 linear constant coefficient difference equation이 있다. y(n) - (1/2)y(n-1) = x(n) + (1/2)x(n-1) ...Unit Impulse Response 계산 (…) 주어진 시스템은 다음과 같은 선형 상수 계수 차분 방정식으로 표현됩니다: $$ y(n) - \frac{1}{2}y(n-1) = x(n) + \frac{1}{2}x(n-1) $$ (…) 위 방정식에 $$x(n) = \d...68677
1368MathematicsMAT_01368직육면체 $B={(x,y,z)\vert0\le x\le 1, -1\le y\le 2, 0\le z\le 3}$일 때 삼중적분 $\iiint_{{B}}xyz^2dV$를 계산해줘.직육면체 ( B = {(x,y,z) \mid 0 \le x \le 1, -1 \le y \le 2, 0 \le z \le 3} )에 대한 삼중적분을 계산하려면 다음과 같은 단계를 따릅니다. (…) $$ \iiint_B xyz^2 , dV = \int_{z=0}^{z=3...68631
  • 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 the four STEM categories and assigned a category-prefixed identifier so that domain balance is auditable at a glance. Problems are weighted toward those that require a derivation or an applied judgment rather than recall of a formula, and both registers are collected deliberately: formally phrased problems, and the conversational phrasing an engineer or student actually uses when asking.

Answers are drafted as complete workings — restate the problem, state the governing equation or principle, define each symbol, substitute the given values, and carry the arithmetic through to the result. Where the answer is a procedure rather than a value, the same shape applies, with runnable code included when the question asks for an implementation. Each answer is then localized into the eight target languages, with LaTeX expressions, symbols, and established English technical terms carried across intact.

Validation

Annotators holding a STEM major verify each answer by re-deriving the result independently rather than reading the provided working. Numeric answers are checked for unit correctness and significant figures, and answers that arrive at the right value through invalid reasoning are rewritten rather than accepted, since a correct number with a broken derivation is worse than useless as training signal.

Localized versions are then checked back against the Korean source to confirm that equations, symbols, and variable names survived unchanged and that no step was dropped in translation. Terminology is cross-checked against the per-language term list so that a single concept is not rendered three different ways across the set. 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 step.

Flitto Curation Data Flitto Curation Data