Back to Gap Map Model components / Training & synthetic datasets

NuminaMath

Project Numina
open / Overall score: 4.0(strong)

NuminaMath is a family of math post-training datasets from Project Numina. NuminaMath-CoT pairs about 860,000 problems, from Chinese high-school exercises to olympiad competitions, with chain-of-thought solutions; NuminaMath-TIR adds about 72,000 solutions that reason by writing and running Python code; and NuminaMath 1.5 revises the collection to about 900,000 problems with answer, problem-type and question-type metadata. Models trained on it won the first AI Mathematical Olympiad Progress Prize.

One entry covers the CoT, TIR and 1.5 releases.

Openness

5 high confidence
5.0
license
apache-2.0 (card metadata and licensing section of all three releases)
access
public (ungated on the Hub)
dataset_card
present (cards give the sources, per-source counts, processing steps and splits)

All three releases are Apache-2.0, download without a gate, and come with cards that list where the problems came from and how they were processed.

Adoption

4 high confidence
4.0

Hugging Face downloads of the three releases over the trailing month, most of them for NuminaMath-CoT. A download count does not say who downloaded, and training pipelines can fetch a dataset repeatedly.

Capability

4 medium confidence
4.0

Fine-tuned on the CoT and then the TIR set, NuminaMath 7B TIR won the first AI Mathematical Olympiad Progress Prize. No controlled comparison against other math instruction sets has been published, so the evidence for its value is that one result rather than matched ablations.

Verified 2026-11-12