ShadenA/MathNet

Quick Start · Overview · Tasks · Comparison · Dataset Stats · Data Sources · Pipeline · Schema · License · Citation

This is MathNet v0. A larger version v1 will be uploaded by Friday, April 24, 2026 (more countires, problems and richer metadata). Schema is stable but field values may be revised in v1.

Quick start

from datasets import load_dataset

# Default: all problems
ds = load_dataset("ShadenA/MathNet", split="train")

# Or a specific country / competition-body config
arg  = load_dataset("ShadenA/MathNet", "Argentina", split="train")
apmo = load_dataset("ShadenA/MathNet", "Asia_Pacific_Mathematics_Olympiad_APMO", split="train")

row = ds[0]
print(row["competition"], row["country"])
print(row["problem_markdown"])
for img in row["images"]:
    img.show()  # PIL image — renders inline in the HF viewer

Overview

Mathematical problem solving remains a challenging test of reasoning for large language and multimodal models, yet existing benchmarks are limited in size, language coverage, and task diversity. We introduce MathNet, a high-quality, large-scale, multimodal, and multilingual dataset of Olympiad-level math problems together with a benchmark for evaluating mathematical reasoning in generative models and mathematical retrieval in embedding-based systems.

MathNet spans 47 countries, 17 languages, and two decades of competitions, comprising 30,676 expert-authored problems with solutions across diverse domains. Alongside the core dataset, we construct a retrieval benchmark of mathematically equivalent and structurally similar problem pairs curated by human experts.

---

Three benchmark tasks

	Task	What it measures
I	Problem Solving	Generative models on Olympiad problems, graded against expert solutions
II	Math-Aware Retrieval	Embedding models' ability to retrieve mathematically equivalent / structurally similar problems
III	Retrieval-Augmented Problem Solving	How retrieval quality affects reasoning when similar problems are given as context

Even state-of-the-art reasoners remain challenged: 78.4% (Gemini-3.1-Pro) and 69.3% (GPT-5) on MathNet-Solve-Test. Embedding models struggle with equivalence retrieval (Recall@1 under 5% for all tested models), and RAG gains are highly sensitive to retrieval quality — expert retrieval lifts DeepSeek-V3.2-Speciale to 97.3% on MathNet-RAG.

How MathNet compares to existing math benchmarks

Benchmark	Size	Languages	Multimodal	Source	Difficulty
GSM8K	8,500	EN	—	Crowdsourced	Grade school
MATH	12,500	EN	—	Competitions/textbooks	High school
MATH-Vision	3,040	EN	✓	Math competitions	High school
OlympiadBench	6,142	EN, ZH	✓	Official websites	Olympiad
OlympicArena	3,233	EN, ZH	✓	Official websites	Olympiad
Omni-Math	4,428	EN	—	AoPS / contest pages	Olympiad
OlymMATH	200	EN, ZH	—	AoPS / official	Olympiad
MathArena	162	EN	✓	Newly released competitions	Olympiad
IMOBench	460	EN	—	IMO & national archives	Olympiad
MathNet (ours)	30,676	17 (EN, ZH, ES, RU, FR, RO, + 11 more)	✓	Official country booklets / international & national contests	Olympiad

Dataset at a glance

What the figure shows. (a) A mix of national, regional, TST, and international competitions. (b) MathNet solutions are substantially longer than those in prior math benchmarks — long-form proofs, not one-line answers. (c) Problems per year — the corpus has grown steadily since the early 2000s. (d) Coverage across geometry, algebra, combinatorics, number theory, and their sub-topics. (e) 74% English, 26% non-English across 17 languages; Portuguese, Spanish, French, Italian, Serbian, Slovenian, German, Chinese, Romanian, Korean, Dutch, Russian, Mongolian, Macedonian, Polish, and Hungarian all appear.

Topic taxonomy (excerpt)

MathNet ships with a curated olympiad-style taxonomy. Top-level domains include:

• Geometry — plane (triangles, quadrilaterals, circles, concurrency/collinearity, transformations, Miquel/Simson/Brocard, geometric inequalities, combinatorial geometry, analytic methods), solid, differential, non-Euclidean
• Algebra — prealgebra, polynomials, inequalities, functional equations, sequences/series, linear algebra, abstract algebra
• Number Theory — divisibility, primes, modular arithmetic, Diophantine equations, quadratic residues, (p)-adic methods
• Combinatorics — counting, graph theory, extremal / pigeonhole, invariants/monovariants, games, coloring, generating functions
• Calculus / Analysis — limits, inequalities, real analysis, combinatorial analysis
• Probability & Statistics — discrete and continuous

Every problem carries a hierarchical topic path (e.g. Geometry > Plane Geometry > Quadrilaterals > Cyclic quadrilaterals) usable for stratified evaluation or curriculum construction.

Data sources

Each year, participating IMO countries contribute original problems for use in their national contests and team selection examinations. MathNet is built from official problem booklets collected from 47 countries spanning 1985–2025 — 1,595 PDF volumes totalling more than 25,000 pages. Unlike prior math benchmarks that rely on community platforms such as AoPS, every problem and solution in MathNet is authored and disseminated by national teams themselves, ensuring expert-level quality, stylistic consistency, and immunity from the noisy or informal annotations that plague crowd-sourced collections.

A meaningful portion of the collection — particularly older national booklets — was physically obtained and scanned by hand by our IMO expert co-authors, who have attended the International Mathematical Olympiad since 2006 and accumulated a personal archive of official competition materials over nearly two decades.

Data pipeline

Extracting aligned problem–solution pairs from a heterogeneous corpus of mathematical documents is non-trivial: some booklets separate problems and solutions into different sections, others interleave them; numbering schemes and naming conventions vary across countries and even within a single document. Regex-based heuristics break down at this scale, so we designed a multi-stage LLM pipeline.

Stage 1 — Document ingestion & segmentation. All booklets are converted to Markdown via dots-ocr, a multilingual document parsing framework designed for both digital typeset PDFs and scanned copies across many languages. Gemini-2.5-Flash then identifies problem and solution segments by outputting only their line numbers, and records authors, hints, remarks, source file, and page numbers for provenance.

Stage 2 — Problem–solution extraction. Given the line segments from Stage 1, GPT-4.1 extracts the corresponding problem and solution in LaTeX-friendly Markdown, together with a surrounding text buffer to handle cases where content spans across context boundaries.

Stage 3 — Extraction verification. Each extracted pair passes three independent checks before being retained:

1. Rule-based similarity check — text similarity between the extraction and original OCR output ensures the LLM made only formatting changes and introduced no hallucinated content.
2. GPT-4.1-as-judge — GPT-4.1 compares page screenshots against the extracted pair to catch OCR errors, incorrect figure associations, and incomplete solutions.
3. Human expert review — low-confidence cases are manually reviewed by annotators. A pair is retained only if all three mechanisms agree.

Provenance (source booklet, authors where given) is preserved on every problem.

What v0 contains

This is the v0 drop of MathNet — the first complete public release:

• 27,817 problems across 58 country / regional-body configs
• 5,148 problems with figures, totalling 7,541 images embedded inline as HF Image() features (they render in the viewer and decode to PIL on load)
• All image references in problem and solution markdown are rewritten to the uniform ![](attached_image_N.png) convention and the corresponding bytes ship in the images column in the same order
• Default all config opens with a curated head of ~120 country-diverse, figure-rich problems so the dataset viewer preview is visually representative; the remainder of all is shuffled

A more refined v1 — larger, with improved extraction, deduplication, and metadata — will be uploaded by Friday, April 24, 2026.

Schema

Column	Type	Notes
id	string	Short 4-char base36 identifier, stable across rebuilds
country	string	Country or regional body of origin
competition	string	e.g. IMO 2023, Cono Sur Mathematical Olympiad
problem_markdown	string	Problem statement (Markdown + LaTeX) with  refs
solutions_markdown	list<string>	Official / provided solutions, one entry per solution
topics_flat	list<string>	Hierarchical topic paths joined as A > B > C
language	string	Source booklet language
booklet_source	string	Upstream collection label
images	list<Image>	Inlined figure bytes, decoded to PIL; positions align with attached_image_N.png refs in the markdown
problem_type	string|null	proof only, answer only, proof and answer — LLM-assisted
final_answer	string|null	LLM-extracted final answer for answer-bearing problems

problem_type and final_answer are LLM-assisted and not fully human-audited in v0. Treat them as convenience annotations, not ground truth.

Configs / splits

One config per country or regional body (58 total) plus a default all config unioning everything. Each config has a single train split — this is the public v0 release, not the train/test partitioning of MathNet-Solve (which is train: 23,776, test: 6,400, test-hard: 500 in the paper release).

Intended uses & limitations

Good for. Olympiad-level reasoning evaluation, multilingual math evaluation, figure-grounded multimodal math, topic-stratified analysis, retrieval benchmarks over mathematical structure, and RL training — the large pool of expert-written solutions provides dense rewards for verifiable-answer problems, while the math-aware similarity pairs open a new axis: rewarding a model for retrieving a structurally equivalent problem is a natural, automatically verifiable signal that does not require a closed-form answer.

Caveats.

• Not contamination-clean. Olympiad problems are indexed widely; assume leakage when evaluating pretrained models.
• v0 field values may be revised in v1 (improved extraction / dedup / metadata).
• LLM-assisted metadata is imperfect.

License

With the kind support of IMO President Gregor Dolinar, we reached out to the leaders of all participating countries and obtained their permission to share this dataset publicly. Where a country or contest organization asserts its own copyright, that copyright is retained and takes precedence — see competition, country, and booklet_source on each row. For all remaining problems where no explicit copyright was asserted, the dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0).

In short: use freely, cite the paper, and respect any explicit rights claimed by the original national team.

If you are a rightsholder with a concern, please open an issue or email shaden@mit.edu.

Citation

@inproceedings{alshammari2026mathnet,
  title     = {MathNet: A Global Multimodal Benchmark for Mathematical
               Reasoning and Retrieval},
  author    = {Alshammari, Shaden and Wen, Kevin and Zainal, Abrar and
               Hamilton, Mark and Safaei, Navid and Albarakati, Sultan and
               Freeman, William T. and Torralba, Antonio},
  booktitle = {International Conference on Learning Representations},
  year      = {2026},
  url       = {https://mathnet.mit.edu}
}

Links

• 🌐 Website & paper: https://mathnet.mit.edu
• 🔭 Browse all 30K problems: https://mathnet.mit.edu/explorer.html
• ✉️ Contact: shaden@mit.edu

_{© 2026 Massachusetts Institute of Technology · MathNet · ICLR 2026}