Nemotron-Cascade-RL-Math: 14,476 Math Problems for Reinforcement Learning

Nemotron-Cascade-RL-Math is a dataset of 14,476 math problems and short answers, created by NVIDIA and last updated on December 16, 2025. It is designed for reinforcement learning in mathematical reasoning, aggregating and decontaminating data from sources like OpenMathReasoning, NuminaMath-CoT, DeepScaleR, and AceReason-Math.

Use Cases

• Training reinforcement learning agents for math problem-solving based on the described problem-answer pairs.
• Benchmarking model performance on mathematical reasoning tasks based on the aggregated and filtered data sources.
• Fine-tuning language models for educational applications based on the high-quality math problems.
• Studying data contamination effects in math benchmarks based on the described 9-gram overlap filtering.

Strengths

• Contains 14,476 math problems and short answers, providing a substantial corpus.
• Aggregates data from four named sources: OpenMathReasoning, NuminaMath-CoT, DeepScaleR, and AceReason-Math.
• Includes a data decontamination process filtering samples with 9-gram overlap against test benchmarks.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but file formats, sample data, and license information are unknown.
• Data may reflect source bias inherent to the aggregated datasets from Hugging Face.

Provenance

Source

NVIDIA

Collection Method

Aggregated from multiple open data sources, followed by decontamination and filtering.

Time Range

null

Freshness

Last updated 2025-12-16 03:31:33; freshness should be verified.

Geography

null