OpenR1-Math-220k: 220,000 Math Problems with DeepSeek R1 Reasoning Traces

OpenR1-Math-220k contains 220,000 mathematical problems paired with multiple reasoning traces generated by DeepSeek R1, released by the open-r1 project in February 2025. Each problem includes two to four distinct reasoning paths derived from NuminaMath 1.5, with correctness verified through automated tools or LLM-based judging.

Use Cases

• Fine-tuning language models using the reasoning traces to improve step-by-step problem solving
• Training reward models by comparing the 2-4 different reasoning paths provided for each problem
• Benchmarking mathematical reasoning capabilities against the verified correct answers

Strengths

• 220,000 math problems
• 2-4 reasoning traces per problem
• Verified correctness for all samples
• Includes LLM-as-a-judge verification for 12% of data

Limitations

• Synthetic reasoning traces generated by DeepSeek R1
• Potential judge bias from Llama-3.3-70B-Instruct
• Domain limited strictly to mathematics

Provenance

Source

open-r1 (based on NuminaMath 1.5)

Collection Method

Synthetic generation via DeepSeek R1 and verification via Math Verify and Llama-3.3-70B-Instruct

Freshness

Last updated February 2025