NuminaMath
Project NuminaNuminaMath 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- 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.
- https://huggingface.co/api/datasets/project-numina/NuminaMath-1.5 recorded 2026-11-12
API JSON: "gated":false; card license apache-2.0.
- https://huggingface.co/api/datasets/project-numina/NuminaMath-CoT recorded 2026-11-12
API JSON: "gated":false; card license apache-2.0.
- https://huggingface.co/api/datasets/project-numina/NuminaMath-TIR recorded 2026-11-12
API JSON: "gated":false; card license apache-2.0.
- https://huggingface.co/datasets/project-numina/NuminaMath-1.5/raw/main/README.md recorded 2026-11-12
"This is the second iteration of the popular NuminaMath dataset", "approximately 900k competition-level math problems" with answer, problem_type and question_type metadata; "The dataset is available under the Apache License, Version 2.0".
- https://huggingface.co/datasets/project-numina/NuminaMath-CoT/raw/main/README.md recorded 2026-11-12
Card metadata "license: apache-2.0"; "Approximately 860k math problems, where each solution is formatted in a Chain of Thought (CoT) manner"; a source breakdown from aops_forum to synthetic_math totalling 859,608; processing from OCR to translation and answer formatting.
- https://huggingface.co/datasets/project-numina/NuminaMath-TIR/raw/main/README.md recorded 2026-11-12
Card metadata "license: apache-2.0"; "we selected approximately 70k problems from the NuminaMath-CoT dataset" and generated tool-integrated reasoning paths; train split 72,441.
Adoption
4 high confidenceHugging 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.
- https://huggingface.co/api/datasets/project-numina/NuminaMath-1.5 recorded 2026-11-12
"downloads":28585 for project-numina/NuminaMath-1.5.
- https://huggingface.co/api/datasets/project-numina/NuminaMath-CoT recorded 2026-11-12
"downloads":138257 for project-numina/NuminaMath-CoT.
- https://huggingface.co/api/datasets/project-numina/NuminaMath-TIR recorded 2026-11-12
"downloads":3733 for project-numina/NuminaMath-TIR.
Capability
4 medium confidenceFine-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.
- https://huggingface.co/blog/winning-aimo-progress-prize recorded 2026-11-12
"our model — NuminaMath 7B TIR — was the winner and managed to solve 29 out of 50 problems on the private test set"; the training recipe section "Good data is all you need" with Chain of Thought and Tool-integrated reasoning stages.
- https://raw.githubusercontent.com/project-numina/aimo-progress-prize/main/README.md recorded 2026-11-12
"our winning solution to the AI Mathematical Olympiad - Progress Prize 1", consisting of a fine-tuning recipe, "Two high-quality training datasets of ~1M math problems and solutions", SC-TIR decoding and validation sets.
Verified 2026-11-12