Dataset Description

The dataset was updated on April 27th, 2026 to fix data formatting issues!

Nemotron-Math-v3 is a large-scale mathematical reasoning dataset containing model-generated reasoning trajectories produced both with and without Python Tool-Integrated Reasoning (TIR). Chain-of-thought (CoT) solutions are generated using DeepSeek-V3.2-Speciale, while Python Tool-Integrated Reasoning (TIR) solutions are generated using DeepSeek-V3.2.

The problems in this dataset are sourced from nvidia/Nemotron-Math-v2, which contains high-quality mathematical problems derived from the Art of Problem Solving (AoPS) community and Math StackExchange/MathOverflow forums. Each problem is solved multiple times under different reasoning regimes (including tool use vs. no tool use), and final answers are verified against the reference answers in nvidia/Nemotron-Math-v2. Only solutions whose final answers match the verified reference are included, resulting in a challenging, clean, and high-quality dataset suitable for training and evaluating mathematical reasoning systems.

All components of the pipeline, including data generation, are implemented using NeMo-Skills.
For detailed information, please refer to the NeMo-Skills documentation in the repository.

This dataset is ready for commercial use.

Dataset Owner(s):

NVIDIA Corporation

Dataset Creation Date:

Created on: Jan 23, 2026
Last Modified on: Apr 27, 2026

License/Terms of Use:

This dataset is governed by the Creative Commons Attribution 4.0 International License (CC BY 4.0) for AoPS-derived samples, while StackExchange-Math-derived samples are governed by the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

Intended Usage

This dataset is intended for:

• Training LLMs to perform structured mathematical reasoning
• Studying tool-augmented reasoning vs. pure language reasoning
• Building long-context or multi-trajectory reasoning systems
• Evaluating LLM reasoning robustness and solution diversity
• Research on reasoning modes, error patterns, and verification pipelines

Dataset Composition and Generation

Problem Sources

The problems are sourced from nvidia/Nemotron-Math-v2, which extracted and filtered problems from:

AoPS Subset

This subset consists of mathematical problems originally derived from the Art of Problem Solving (AoPS) community.

Characteristics:

• Competition-style problems across algebra, geometry, number theory, and combinatorics
• Proof-style questions removed to ensure answer verifiability
• Difficulty filtering applied to remove trivially solvable problems

StackExchange-Math Subset

This subset consists of mathematical problems collected from Math StackExchange and MathOverflow.

Characteristics:

• Wide range of topics from undergraduate-level to research-oriented mathematics
• Proof-style questions filtered via LLM classifier
• Decontaminated to avoid overlap with public benchmarks

Dataset Fields

The dataset contains the following fields:

• uuid: Unique identifier for the sample.
• problem: Problem statement derived from OpenMathReasoning, Math StackExchange, and MathOverflow.
• messages: Standardized multi-turn message sequence for LLM training. Depending on the sample, this can include system, user, assistant, and tool roles.
• expected_answer: Verified target answer for the sample.
• changed_answer_to_majority: Boolean indicator showing whether the original extracted answer was replaced by a majority-vote answer from generated solutions.
• data_source: Source of the problem, either AoPS or StackExchange-Math.
• license: Per-sample license label. AoPS-derived samples use cc-by-4.0, while StackExchange-derived samples use cc-by-sa-4.0.
• tool_usage: Reasoning regime label indicating whether the sample was generated with Python TIR or without Python TIR.
• tools: Tool definition list. This is empty for non-tool samples and contains the Python tool specification for TIR samples.
• url: Hyperlink to the original question.
• user_url: Hyperlink to the original author profile.
• user_name: Original author username.
• used_in: Reserved list field for downstream usage annotations.

Dataset Characterization

Data Collection Method

• Hybrid: Human, Synthetic, Automated

Labeling Method

• Hybrid: Human, Synthetic, Automated

Dataset Format

Modality: Text
Format: JSONL
Structure: JSONL records with text, messages, labels, provenance, and tool annotations

Dataset Quantification

Total Disk Size: ~144 GB

Reference(s):

• nvidia/Nemotron-Math-v2
• nvidia/OpenMathReasoning
• NeMo-Skills
• Link to paper.

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here