NuminaMath

NuminaMath-CoT is a large bank of competition mathematics problems, each paired with a worked, step-by-step solution written in a chain-of-thought style. It was assembled by Project Numina (published under the AI-MO organisation on Hugging Face) and released in 2024 as part of their work on competition-level maths models. The training split holds roughly 860,000 problem-solution pairs (about 859,608 rows), spanning Chinese high school exercises through to US and international olympiad problems, so the difficulty range is genuinely broad. Subsets are tagged by origin, including cn_k12, olympiads, orca_math, synthetic_math, aops_forum, amc_aime, gsm8k and math, which lets you filter by level or provenance.

You access it directly on Hugging Face as Parquet, streamable through the datasets library with no account or application required, which is why the access tier is open. The problems were largely OCR'd from online exam PDFs and discussion forums, then segmented into problem-solution pairs, translated into English and realigned into a consistent CoT format.

For RAG it is a ready supply of worked examples and final answers to retrieve against for maths and reasoning tasks. For an AI tutor it is more directly useful: the origin tags support rough level sequencing (school exercises before olympiad), and each item is a self-contained practice problem with a modelled solution you can show, hint from, or grade against, which suits practice generation and mastery checking.

The dataset wrapper is licensed Apache 2.0, so commercial use is permitted with the usual notice retention. The important caveat is provenance: individual problems were scraped and OCR'd from exam papers and forums, so per-item rights are mixed and the Apache licence does not cover the underlying content. Treat it accordingly in a commercial product.

It complements the pretraining-scale megamath and the question-shaped stackmathqa, sits alongside amps, openthoughts3 and naturalproofs as reasoning-oriented maths data, and pairs well with structured curricula such as openstax, khan-academy, ck-12, siyavula and common-core when you need pedagogy rather than raw problems.