7 Essential Math Reasoning Datasets for AI: From Arithmetic to Visual Geometry

As large‑model capabilities advance, mathematical reasoning has become a frontier challenge for artificial intelligence, requiring models to grasp not only surface meanings but also underlying logical structures; consequently, high‑quality, structured datasets are crucial for training and evaluating such models.

We‑Math2.0‑Standard (≈369.86 MB) is a visual‑math benchmark released in 2025 by Beijing University of Posts and Telecommunications, Tencent, and Tsinghua University. The associated paper "WE‑MATH 2.0: A Versatile MathBook System for Incentivizing Visual Mathematical Reasoning" (https://hyper.ai/en/papers/2508.10433) describes a unified label space of 1,819 knowledge principles, with each sample containing multiple images and multiple questions per image, together with explicit principle annotations and standard answers.

NuminaMath‑LEAN (≈65.06 MB) was jointly released in 2025 by Numina and the Kimi Team. Its paper "Kimina‑Prover Preview: Towards Large Formal Reasoning Models with Reinforcement Learning" (https://hyper.ai/en/papers/2504.11354) provides 100 k competition problems from IMO, USAMO, etc., each annotated with problem type, answer, source, formal statement, proof, and reinforcement‑learning training logs.

T‑Wix Russian SFT (≈1.43 GB) originates from a Russian effort and is described in "From Quantity to Quality: Boosting LLM Performance with Self‑Guided Data Selection for Instruction Tuning" (https://arxiv.org/abs/2308.12032). It contains 499 598 Russian samples across math, science, programming, commonsense, instruction following, and role‑play, of which 30 984 are dedicated reasoning examples with detailed reasoning trajectories.

Nemotron‑Math‑HumanReasoning (≈639.91 KB) was released by NVIDIA in 2025. The accompanying paper "The Challenge of Teaching Reasoning to LLMs Without RL or Distillation" (https://arxiv.org/abs/2507.09850) offers 50 math questions from the OpenMathReasoning dataset, 200 human‑written solutions, and 50 additional model‑generated answers.

Open‑Omega‑Atom‑1.5M (≈6.6 GB) is a large‑scale math‑and‑science reasoning collection containing about 1.5 million entries, emphasizing concise, high‑quality, step‑by‑step problem‑solution pairs and a mix of mathematics, code, and scientific reasoning.

GSM8K (≈4.92 MB) was released by OpenAI in 2022 (paper: "Training Verifiers to Solve Math Word Problems", https://arxiv.org/abs/2110.14168). It comprises 8.5 k elementary‑level word problems covering algebra, arithmetic, and geometry, with solutions ranging from 2 to 8 steps using basic operations (+, –, ×, ÷).

VCBench (≈86.04 MB) was jointly published by Alibaba and Zhejiang University in 2025. The benchmark contains 1 720 question‑answer pairs and 6 697 images, organized into six domains: time & calendar, space & position, geometry & shape, object & motion, reasoning & observation, and organization & pattern, each testing specific visual‑mathematical reasoning abilities.

Collectively, these datasets provide diverse, well‑curated resources that enable researchers to diagnose, compare, and improve the reasoning capabilities of large language models across arithmetic, symbolic logic, visual mathematics, and geometric analysis.