Nemotron-Math-HumanReasoning: Olympiad-Level Math Solutions

NVIDIA released this dataset in July 2025 containing fewer than 1,000 human-written solutions to complex math problems. It features extended reasoning chains authored by Olympiad-level mathematics students to emulate the step-by-step logic of advanced reasoning models.

Use Cases

• Fine-tuning LLMs using the human-written solutions to improve step-by-step mathematical reasoning
• Evaluating model performance on Olympiad-level math problems
• Analyzing human reasoning patterns in complex problem-solving

Strengths

• Expert-authored solutions by Olympiad-level math students
• Focuses on extended reasoning chains similar to DeepSeek-R1 style
• High-quality human-written data rather than synthetic or distilled outputs

Limitations

• Small dataset size with fewer than 1,000 examples
• Restricted to mathematical reasoning domain
• Non-commercial license (CC BY-NC 4.0) limits usage

Provenance

Source

NVIDIA (Arxiv: 2507.09850)

Collection Method

Annotated (authored by students with Olympiad-level math experience)

Freshness

Last updated July 2025.