OpenMathReasoning: Large-scale math reasoning dataset from NVIDIA
2 min read
This dataset contains hundreds of thousands of unique math problems and millions of chain-of-thought solutions generated using cutting-edge models like DeepSeek-R1 and QwQ-32B. The dataset served as the foundation for a winning solution in the Kaggle AIMO-2 competition and includes problems from AoPS forums. It is ideal for researchers and developers fine-tuning language models to significantly boost their ability to solve complex competition and math problems.
Key features
- Over 306,000 unique math problems with detailed long-form solutions (chain-of-thought and TIR).
- Solutions generated using powerful DeepSeek-R1 and QwQ-32B models.
- Includes GenSelect mode for picking the most promising solutions from multiple candidates.
- Served as the foundation for creating the cutting-edge OpenMath-Nemotron model series.
What you can do
- Fine-tune language models (LLMs) to enhance math and logical reasoning capabilities.
- Train models to use tools (TIR – tool-integrated reasoning) when tackling computational tasks.
- Use the dataset to evaluate and benchmark the math capabilities of custom models.
- Leverage the data to replicate training pipelines similar to the Kaggle AIMO-2 winner.
Example
Example of loading the dataset using the datasets library:
from datasets import load_dataset
dataset = load_dataset("nvidia/OpenMathReasoning")
print(dataset["train"][0])Read more
DetailsVolume: from 1M to 10M compressed records.
Primary sources: AoPS forums and the MATH dataset.
License: CC-BY-4.0.
Metrics: OpenMath-Nemotron models trained on the dataset achieve state-of-the-art results on AIME24, AIME25, and HLE-Math benchmarks.








Comments