Omni-MATH

Omni-MATH is an olympiad-level mathematics benchmark of 4,428 competition problems, each paired with a human-annotated solution, built by Bofei Gao and colleagues and released in 2024 alongside the paper "Omni-MATH: A Universal Olympiad Level Mathematic Benchmark for Large Language Models". Every problem is tagged with one of 33 or more mathematical sub-domains (algebra, number theory, geometry, combinatorics and more) and placed on a scale of roughly ten difficulty levels, so the collection is graded rather than a flat pile of hard questions.

You access it freely through the Hugging Face datasets library (the KbsdJames/Omni-MATH repository), which serves JSON and Parquet with fields for the problem statement, the worked solution, the domain and the difficulty. No account or application wall stands in the way, and a mirror sits on Kaggle, with a companion GitHub repository and project page holding the evaluation code.

For RAG, the value is a clean, structured pool of hard problems and reference solutions: you can embed each problem or solution as a record, retrieve similar worked examples to ground a model's reasoning, or use the domain and difficulty tags as retrieval facets. For an AI tutor the same structure pays off differently. The difficulty scale supports curriculum sequencing and mastery modelling, letting a tutor step a strong learner up level by level, while the sub-domain tags let you target a specific weakness, and the human-written solutions offer worked pedagogy rather than bare answers. Its practical role is to stretch advanced students at the top of the curriculum, well beyond school syllabuses.

The dataset card licences the release under Apache 2.0, which permits commercial use and asks only that you preserve the notice. Treat that as covering the annotations and structure: the underlying problems come from real competitions and olympiads whose original wording may carry its own rights, so check provenance before commercial redistribution of the text.

Among our maths sources it is the elite-competition tier: harder than the school-aligned OpenStax, Khan Academy, CK-12, Siyavula, Common Core and NGSS material, and complementary to the reasoning corpora AMPS, StackMathQA, MegaMath, OpenThoughts3 and NaturalProofs, which lean towards volume, proofs or general practice rather than curated olympiad difficulty.