Nemotron Math Reasoning Dataset for Reinforcement Learning

Nemotron-Cascade-RL-Math is a dataset of 14,476 math problems and short answers for reinforcement learning. It was created by NVIDIA, aggregating and filtering content from sources like OpenMathReasoning and NuminaMath-CoT. The dataset was last updated in December 2025.

Use Cases

• Training reinforcement learning agents on math problem-solving tasks using the provided problems and short answers.
• Benchmarking the mathematical reasoning capabilities of language models on a filtered set of 14,476 problems.
• Fine-tuning models for generating concise, step-by-step solutions to math problems from the aggregated sources.
• Studying data contamination by using the 9-gram overlap filter applied to benchmark test samples.

Strengths

• Contains 14,476 math problems, providing a substantial collection for training.
• Aggregates content from multiple established sources including OpenMathReasoning and NuminaMath-CoT.
• Includes a data decontamination step filtering samples with 9-gram overlap against benchmark tests.

Limitations

• The specific types of math problems, difficulty levels, and answer formats are not detailed.
• Unknown if the dataset includes step-by-step reasoning chains or only final short answers.
• The total size, row count, and column structure are not provided in the available description.

Provenance

Source

Aggregated from OpenMathReasoning, NuminaMath-CoT, DeepScaleR, and AceReason-Math.

Collection Method

Data was aggregated from source datasets, decontaminated, and filtered for 9-gram overlap with test benchmarks.

Freshness

Last updated on 2025-12-16.