You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Multi-Math

Multi-Math is a multilingual collection of synthetic math problems with worked solutions, released in English, Italian, French, German and Spanish. It contains about 87,000 problem–solution pairs in total, generated from a curated set of seed problems collected from public web sources.

The dataset is a supervised fine-tuning (SFT) resource for mathematical reasoning: each sample pairs a problem statement with its complete solution, both written in LaTeX, and is annotated with its language and difficulty level.

Source data

Seed problems were collected from public web sources and cover a wide range of mathematical topics. The documents were normalized to PDF, converted to one image per page and then cleaned manually: overly long, chaotic or questionable material was removed or split, and every problem was paired with its corresponding solution.

Because the licensing of the collected material is heterogeneous and in many cases unknown or restrictive, the seeds are not redistributed. The published dataset contains only newly generated problems and solutions.

How the dataset was built

  1. Seed extraction: Qwen3-VL-235B-A22B-Instruct was used to extract problem–solution pairs from the page images as structured output with LaTeX text. When a problem had more than one solution, a separate pair was created for each solution, and every pair was annotated with whether its solution depends on a figure.
  2. Difficulty assessment: each pair was labelled Easy, Medium or Hard, from elementary-school level up to university level. The Hard pairs were discarded, since the generative models could not produce coherent problems and solutions at that level of complexity.
  3. Generation: the final problems were generated with Qwen3.5-122B-A10B with thinking enabled, producing three new problems for each seed. Problems that require an image and problems of Hard difficulty were discarded.

Dataset structure

Configurations

Config Language Entries Share
en English 53,318 61.3%
it Italian 16,679 19.2%
fr French 8,618 9.9%
de German 6,911 7.9%
es Spanish 1,468 1.7%
Total 86,994 100%

The configurations are not balanced across languages: the number of samples reflects the number of seed problems available for each language.

Fields

Field Description
ID Unique identifier of the sample
Problem Problem statement, written in LaTeX
Solution Complete solution of the problem, written in LaTeX
Language Language of the sample: en, de, es, fr or it
Difficulty Difficulty level of the problem: Easy or Medium

Usage

from datasets import load_dataset

ds = load_dataset("VillanovaAI/multi-math", "en", split="train")

sample = ds[0]
print(sample["Language"], sample["Difficulty"])
print(sample["Problem"])
print(sample["Solution"])

Limitations

  • The problems and solutions are model-generated and may contain incorrect statements, wrong computations or inconsistent LaTeX; the solutions should be verified before being used for training or evaluation.
  • Only Easy and Medium problems are included, so advanced, competition-level and university-level mathematics is not represented.
  • The dataset contains no diagram-based or geometry-from-figure tasks.
  • The language configurations are unbalanced, with Spanish being much smaller than the others.

Licensing and attribution

The released dataset is distributed under ODC-BY. The seed problems were collected from public sources under heterogeneous and often unknown or restrictive licenses, so they are not redistributed and no reference to them is included in the published data.

Downloads last month
60