OpenMathReasoning

OpenMathReasoning is a large mathematical reasoning corpus released by NVIDIA in 2025, and the data foundation of their winning submission to the AIMO-2 Kaggle competition. It pairs roughly 306,000 unique competition problems, sourced from the Art of Problem Solving (AoPS) community forums, with model-generated solutions, namely about 3.2 million long chain-of-thought (CoT) traces, about 1.7 million tool-integrated reasoning (TIR) traces that interleave natural language with executed code, and around 566,000 generative solution selection (GenSelect) samples. A further 193,000 problems are included without solutions. Problem statements were refined with Qwen2.5-32B-Instruct, and the solutions were produced by DeepSeek-R1 and QwQ-32B.

You access it directly on Hugging Face as Parquet, streamable through the datasets library with no account or application, which is why the access tier is open. The full download is around 49.5 GB across the cot, tir, genselect and additional_problems splits.

For AI work its main use is fine-tuning and distilling reasoning models, where the explicit CoT and TIR traces are training targets for teaching step-by-step and code-assisted problem solving; NVIDIA's own OpenMath-Nemotron models were trained on exactly this data. For RAG it is a bank of worked solutions and final answers to retrieve against for maths and reasoning tasks, or to seed few-shot examples. It is best suited to competition-level and olympiad mathematics rather than elementary drill.

The dataset is licensed CC BY 4.0, so commercial use is permitted provided you attribute NVIDIA and cite the accompanying paper (AIMO-2 Winning Solution). Two caveats matter. The underlying problems come from AoPS forums and carry their own terms, so the content licence differs from the permissive wrapper, and every solution is machine-generated by DeepSeek-R1 or QwQ-32B, which some downstream users screen or filter before use.

It sits alongside numinamath, big-math-rl-verified, openthoughts3 and omni-math as reasoning-oriented maths data, and complements the pretraining-scale megamath.