VietAlpha A research lab for Vietnam's mathematical record

Vietnamese-language mathematics datasets

Vietnamese-language mathematics datasets include translated evaluation benchmarks, scraped elementary question banks, examination archives, and large instruction-tuning corpora derived from machine translation.

Each entry gives what the dataset's own paper or card states, and says where that documentation stops. These resources are the machine-readable layer of the Vietnamese mathematical record; the institutions that study the subject are mapped separately in mathematics research in Vietnam.

The four kinds of published data.

The published Vietnamese mathematics data falls into four kinds: translations of English-origin benchmarks, elementary collections scraped from educational sites, examination-derived evaluation sets, and large instruction dumps with little documentation. Some efforts adapt established English items through machine or human translation, while other projects compile questions directly from domestic online educational platforms or official academic tests. Large instruction dumps appear from automated generation pipelines or general translation efforts, often arriving with sparse documentation regarding their exact provenance. Several of the cards carry no README at all, so the origin and method of those releases cannot be read from them.

The datasets.

The order below is arbitrary. This list carries no ranking and no judgement of quality.

  • VNHSGE

    Examination-derived evaluation set

    VNHSGE was generated from the Vietnamese National High School Graduation Examination and comparable tests.1 It covers nine subjects and holds over 19,000 multiple-choice questions, and includes 300 literary essays.1 Evaluating models on it, its authors wrote that they "still have space to grow, though, especially in the areas of mathematics, physics, chemistry, and biology".1

  • ViMath-Bench

    Elementary collection

    ViMath-Bench is described by its authors as "the first large-scale Vietnamese elementary mathematical dataset".2 After preprocessing it contains 8.4K samples, divided into 7.1K training and 1.3K test samples.2 Its items were collected from three Vietnamese online educational platforms, Tailieumoi, Hamchoi and Vietjack, and the authors "specifically targeted multiple-choice questions (MCQs) from grades 3 to 5".2

  • ViMath-InstructCode

    Instruction-tuning set derived from ViMath-Bench

    ViMath-InstructCode is a separate, augmented training set of 14K samples, produced from ViMath-Bench using strong teacher models.2 The dataset expands upon the raw collected problems by introducing generated code solutions and step-by-step reasoning paths.2

  • stair-lab/MATH_vi

    Translation of an English-origin benchmark

    The HuggingFace dataset stair-lab/MATH_vi is a Vietnamese translation.3 Its card names the original version as lighteval/MATH and links translation source code.3

  • ura-hcmut/Vietnamese-MATH

    Parallel English and Vietnamese release

    ura-hcmut/Vietnamese-MATH is a parallel English and Vietnamese release with seven subject subsets: algebra 2.93k rows, counting and probability 1.25k rows, geometry 1.35k rows, intermediate algebra 2.2k rows, number theory 1.41k rows, prealgebra 2.08k rows, precalculus 1.29k rows.4 Its card does not identify its original benchmark and does not describe its translation method.4

  • 5CD-AI instruction sets

    Machine-translated instruction dumps

    In 2024, 5CD-AI published Vietnamese machine-translated mathematics instruction sets: MetaMathQA at 395k rows, Microsoft Orca Math word problems at 200k rows, and NVIDIA OpenMathInstruct at 50k rows.5 The 395k card is labelled English and Vietnamese and its README is empty, so nothing about method or curriculum can be read from it.5

  • DHMATH/SFT_Data

    Instruction dump

    DHMATH published a Vietnamese mathematics supervised fine-tuning release of 168k rows with an elementary subset of 29.5k rows.6 Its README is empty, so its origin and method cannot be read from the card.6

  • BenchMAX_Math

    Translation of an English-origin benchmark

    BenchMAX_Math sources from MGSM and extends it with six further languages including Vietnamese.7 Its card states the data "is first translated by Google Translate, and then post-editing by native speakers".7

  • Vi-MQA

    Examination-derived evaluation set

    Vi-MQA, published with VMLU, collected standardized examinations from seven primary subjects in the Vietnamese general education curriculum for the elementary, middle and high school levels, with mock exams drawn from freely available online materials.8 Its mathematics counts are Elementary Mathematics 200, Middle School Mathematics 119, and High School Mathematics 163.8

  • vi-gsm8k-agentic

    Native Vietnamese synthetic set

    vi-gsm8k-agentic is 1,465 rows of native Vietnamese grade-school word problems with code-verified numeric answers and step-by-step solutions.9 Its card says it was generated by an agentic self-instruct pipeline and is not machine-translated.9

  • Vietnamese-Entrance-Exam

    Examination-derived evaluation set

    Vietnamese-Entrance-Exam is 432 problems derived from Vietnamese university entrance examinations, of which 243 are mathematics.10 Its card states the original Vietnamese problems were translated to English.10

  • LuminarAI textbook mathematics

    Textbook-oriented instruction dump

    LuminarAI published a Vietnamese middle and high school textbook-oriented mathematics release of 306k rows, whose card says it focuses on core mathematics topics within Vietnam's general education curriculum.11 The card states no per-item curriculum labelling.11

  • UIT-DODV

    Document-image dataset

    UIT-DODV is described as the first Vietnamese document image dataset, with 2,394 images and four classes: Table, Figure, Caption, Formula.12 Its task is object detection, with bounding boxes and labels as the output, so Formula marks a region of a page and is not transcribed mathematics.12

  • Vietnamese mathematical-expression images

    Mathematical-expression recognition set

    A published dataset holds 6,000 Vietnamese camera-captured document images containing mathematical expressions.13 These images provide both position information and LaTeX strings for the expressions, intended for evaluating recognition algorithms.13

Curriculum alignment.

Searching in July 2026, across English and Vietnamese sources and HuggingFace, VietAlpha found no openly published Vietnamese mathematics dataset whose paper or dataset card states alignment to the GDPT 2018 mathematics programme, or labelling at the level of a curriculum unit. A translated benchmark inherits the curriculum of the system it came from, which is a different thing from the mathematics curriculum taught in Vietnamese schools. In the SeaLLMs evaluation, models scored lower on Vietnamese translations of GSM8K and MATH than on the English originals.14 For SeaLLM-7B-v2.5 the English and Vietnamese figures are GSM8K 78.5 against 72.3, and MATH 34.9 against 30.2.14

Several resources, including stair-lab/MATH_vi, ura-hcmut/Vietnamese-MATH, BenchMAX_Math, and the 5CD-AI releases, consist of translations or machine translations of English-origin material. ViMath-Bench collects items from three Vietnamese educational websites and targets grades 3 to 5. VNHSGE, Vi-MQA, and Vietnamese-Entrance-Exam are drawn from Vietnamese examinations. DHMATH/SFT_Data and the LuminarAI release are large, and their cards do not describe how the items were made.

Two of them state a relationship to the Vietnamese general education curriculum in general terms. Vi-MQA says its examinations come from seven subjects of that curriculum, and the LuminarAI card says it focuses on core mathematics topics within it. Neither of those two states which unit of the programme an individual item belongs to. A dataset can be Vietnamese in language, and Vietnamese in origin, and still carry no label saying which part of the taught programme an item tests.

Mathematical text recognition.

Mathematical text recognition for Vietnamese spans document layout detection and expression transcription. UIT-DODV labels document images with bounding boxes for tables, figures, captions, and formulas, but does not transcribe the mathematical symbols inside those regions.12 A separate dataset provides 6,000 camera-captured images with position information and LaTeX strings for evaluating recognition algorithms.13 A 2025 survey of Vietnamese document analysis and recognition states that "there is a notable scarcity of large-scale annotated datasets for Vietnamese document images" and that this limits development and fine-tuning of models for Vietnamese OCR.15 Its stated scope is Vietnamese document analysis generally, and it is not a survey of mathematical-expression recognition.15 General tools like Nougat convert scientific documents to markup to address the loss of semantic information in PDF "particularly for mathematical expressions".16 Other tools like pix2tex take an image of a mathematical formula and return the corresponding LaTeX code, and its own documentation tells users to "Always double check the result carefully".17

How this lab builds data.

The research lab published its dataset work in rounds. In the first round of work, a strict schema contract resolved contamination between question and solution boundaries, and confirmed that parsing Vietnamese mathematics PDFs is feasible. The second round was a feasibility study of whether structured Vietnamese mathematics records could be produced from raw PDFs via extraction on local hardware.

In the third round, Code-Flow-v1 is the first end-to-end prototype, and it connected PDF extraction, Markdown conversion, JSON structuring and validation. Cadmus converts Vietnamese mathematical source material into structured, verified, curriculum-tagged records through a resumable, auditable eleven-stage production pipeline.

Normalizing a document before chunking it destroys mathematical tokens, and a dual-view normalization and chunking stage was built to prevent that. The dual-view architecture embeds a normalized view while assembling the output from the original text, and it preserved 100% of mathematical tokens, against 0.03% for a normalization-only baseline. A precision-first post-OCR corrector for Vietnamese mathematics answer keys is deterministic and resource-light, and it reached 99.913% precision across 2,288 automatic edits, with zero catastrophic errors and zero source-file mutations.

A separate note explains how neural networks relate to Vietnamese mathematical writing and works through the forward pass, layer Jacobians, backpropagation and the second-order test.

By its own account, VietAlpha built the first curriculum-aligned mathematics dataset to the standards of Vietnam's Ministry of Education and Training.

Sources.

  1. VNHSGE dataset: Vietnamese High School National Graduation Examination Dataset for Large Language Models (arXiv:2305.12199)

    The dataset's origin in the national graduation examination, its nine subjects, the count of over 19,000 multiple-choice questions and 300 literary essays, and the authors' quoted assessment of model performance.

  2. Nguyen, Nguyen, Nguyen and Bui, A Novel Instruction Tuning Method for Vietnamese Mathematical Reasoning using Trainable Open-Source Large Language Models, CoNLL 2024, pp. 259-268

    The description of ViMath-Bench as the first large-scale Vietnamese elementary mathematical dataset, its 8.4K samples split 7.1K and 1.3K, its three source platforms and grade range, and ViMath-InstructCode as a separate 14K augmented set.

  3. HuggingFace dataset card, stair-lab/MATH_vi

    The card's statement that this is a Vietnamese translation, its naming of lighteval/MATH as the original version, and its link to translation source code.

  4. HuggingFace dataset card, ura-hcmut/Vietnamese-MATH

    The seven subject subsets and their row counts, and the absence from the card of any statement identifying the original benchmark or describing the translation method.

  5. HuggingFace, 5CD-AI Vietnamese translated mathematics instruction datasets

    The three 2024 releases and their row counts, and the empty README on the 395k card.

  6. HuggingFace, DHMATH/SFT_Data

    The 168k rows, the 29.5k elementary subset, and the empty README.

  7. HuggingFace, LLaMAX/BenchMAX_Math

    The sourcing from MGSM, the extension to six further languages including Vietnamese, and the quoted description of the translation and post-editing method.

  8. Vi-MQA and VMLU, ACL 2025

    The collection of standardized examinations across seven subjects and three school levels, the use of freely available online mock exams, and the three mathematics item counts.

  9. HuggingFace, vuongtsc/vi-gsm8k-agentic

    The 1,465 rows, the code-verified numeric answers and step-by-step solutions, and the card's statement that the set was generated by an agentic self-instruct pipeline and is not machine-translated.

  10. HuggingFace, Intelligent-Internet/Vietnamese-Entrance-Exam

    The 432 problems, the 243 mathematics items, and the card's statement that the original Vietnamese problems were translated to English.

  11. HuggingFace, LuminarAI/vietnamese-ms-hs-textbook-math-300K

    The 306k rows and the card's statement of focus on core mathematics topics within Vietnam's general education curriculum, with no per-item curriculum labelling stated.

  12. Parsing Digitized Vietnamese Paper Documents (UIT-DODV)

    The description of UIT-DODV as the first Vietnamese document image dataset, its 2,394 images and four classes, and the statement of the task as object detection with bounding boxes and labels as the output.

  13. Mendeley Data, Vietnamese camera-captured document images containing mathematical expressions

    The 6,000 images, and the provision of both position information and LaTeX strings for the mathematical expressions, for evaluating recognition algorithms.

  14. SeaLLMs, ACL 2024 system demonstrations

    The comparison of GSM8K and MATH scores on the English originals and their Vietnamese translations, and the SeaLLM-7B-v2.5 figures quoted above.

  15. A Survey on Vietnamese Document Analysis and Recognition (arXiv:2506.05061)

    The quoted statement about the scarcity of large-scale annotated datasets for Vietnamese document images, and the survey's own statement of its scope.

  16. Nougat: Neural Optical Understanding for Academic Documents (arXiv:2308.13418)

    The description of the loss of semantic information in PDF, particularly for mathematical expressions, and the conversion of scientific documents to markup.

  17. pix2tex, LaTeX-OCR documentation

    The description of the tool as taking an image of a mathematical formula and returning the corresponding LaTeX code, and its own instruction to check the result carefully.

No published card or paper we found states which part of the Vietnamese curriculum an item belongs to.