Nemotron-Math is a large-scale mathematical reasoning dataset released by NVIDIA in 2026, with the associated paper titled "Nemotron-Math: Efficient Long-Context Distillation of Mathematical Reasoning from Multi-Mode Supervision", designed to provide high-quality data for training and evaluating the mathematical reasoning capabilities of large language models.

The dataset contains 3,638,783 training samples, with a total disk size of approximately 144 GB. The problems in the dataset originate from the AoPS community and the Math StackExchange/MathOverflow forums, and have undergone strict filtering and validation, retaining only high-quality solution trajectories where the final answer matches the reference. The dataset covers chain-of-thought (CoT) solutions in two modes: those using Python tool-integrated reasoning (TIR) and those without tools, making it suitable for research on tool-augmented reasoning, pure language reasoning, and multi-trajectory reasoning systems. Samples originating from AoPS are licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0); samples originating from StackExchange Math are licensed under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

Dataset Composition

• uuid: Identifier for the sample
• problem: Problem statement, sourced from AoPS, Math StackExchange, and MathOverflow
• messages: Standardized multi-turn conversation sequence, including system, user, assistant, and tool roles, used for LLM training
• expected_answer: Verified target answer
• changed_answer_to_majority: Boolean value indicating whether the originally extracted answer was replaced by a majority vote of generated answers
• data_source: Problem source, categorized as either AoPS or StackExchange-Math
• license: License label for each sample; AoPS samples use cc-by-4.0, StackExchange samples use cc-by-sa-4.0
• tool_usage: Reasoning mode label, indicating whether the sample uses Python TIR
• tools: List of tool definitions; empty for non-tool samples, contains Python tool specifications for TIR samples
• url: Hyperlink to the original problem
• user_url: Hyperlink to the original author's profile
• user_name: Username of the original author
• used_in: Reserved field for downstream usage annotation