Skip to content
Dataset By Humans
LLM training data

Vietnamese math datasets for LLMs: options and pitfalls

Translated, scraped, generated or teacher-written: where Vietnamese math reasoning data comes from, how contamination skews scores, and what a good record looks like.

By Dataset By HumansPublished 1 min read

In this article

Short answer: most Vietnamese math data available to LLM teams today is either translated from English benchmarks (such as GSM8K or MATH), scraped from exam and homework sites, or generated by other models. That is fine for quick experiments, but for training reasoning you want problems written in natural Vietnamese with step-by-step solutions checked by teachers, and for evaluation you want a held-out set that has never been published.

Why Vietnamese math is its own problem

  • Language. Vietnamese word problems use phrasing, units and naming conventions that translation flattens or gets wrong.
  • Curriculum. Students learn specific methods and notation for the national high-school exam; a tutor model should explain the way they are taught.
  • Contamination. Problems copied from public exam sites are likely already in the pre-training data of large models, so scores on them overstate real ability.
  • Generated solutions. Chains of thought written by another model copy its mistakes and style.

Where Vietnamese math data comes from

SourceStrengthsWeaknesses
Translated English benchmarksLarge, well knownUnnatural Vietnamese, foreign context
Public exam and homework sitesReal curriculum problemsLicensing unclear, often contaminated
LLM-generated problems and solutionsCheap, unlimitedErrors, repetitive style, terms-of-use questions
Teacher-written, double-checkedAccurate, natural, privateCosts more per record

If you found this page searching for a public set such as VNMath, check its license, its source and whether its solutions were written by people before you train on it.

What a good Vietnamese reasoning record looks like

{
  "id": "vmr-001204",
  "level": "grade-9",
  "topic": "quadratic equations",
  "problem": "Giải phương trình x² − 5x + 6 = 0.",
  "steps": [
    "Tính biệt thức: Δ = (−5)² − 4·1·6 = 1.",
    "Vì Δ > 0, phương trình có hai nghiệm phân biệt.",
    "x₁ = (5 + 1)/2 = 3, x₂ = (5 − 1)/2 = 2."
  ],
  "answer": "x = 2 hoặc x = 3",
  "author_role": "math teacher",
  "reviewed_by_second_teacher": true
}

Step-level structure lets you train process reward models that judge each step, not only the final answer.

Training vs evaluation sets

Use separate data. A training set can be large and is licensed non-exclusively. An evaluation set should be smaller, exclusive to you and never published, so it stays out of every model's training data and your benchmark keeps meaning something.

Getting teacher-written data

Our Vietnamese math reasoning dataset is written by Vietnamese math teachers and checked by a second teacher, from secondary school to university-entrance level. We also build human-authored SFT datasets for other subjects and domains.

vietnamese math datasetreasoning datasetchain of thoughtllm evaluation