NuminaMath CoT: 860,000 Math Problems with Chain of Thought Solutions

AI-MO released this collection of approximately 860,000 math problems with step-by-step Chain of Thought (CoT) solutions in late 2024. The data spans Chinese high school exercises to international Olympiad competitions, sourced from PDFs and online forums via OCR.

Use Cases

• Fine-tuning LLMs for mathematical reasoning using the Chain of Thought (CoT) solution format
• Benchmarking model accuracy on high-difficulty Olympiad competition problems
• Training text generation models to produce structured, step-by-step logical derivations

Strengths

• 860,000 math problems
• Chain of Thought (CoT) formatting
• Includes elite Olympiad-level competition problems

Limitations

• OCR-induced noise in mathematical symbols and notation
• Potential linguistic artifacts from translated Chinese high school sources

Provenance

Source

AI-MO

Collection Method

OCR from original PDFs and segmentation of mathematics discussion forums

Freshness

Last updated November 2024.

Geography

US, China, and International