Skip to content

Dataset · open-r1

OpenR1-Math-220k: Math Reasoning

open-r1/OpenR1-Math-220k

220k math problems with verified reasoning traces from DeepSeek R1.

What's inside

A math reasoning dataset containing 220k problems with two to four reasoning traces generated by DeepSeek R1 based on NuminaMath 1.5 problems. Most traces are verified by Math Verify, and 12% by the Llama-3.3-70B-Instruct judge. Each problem has at least one trace with the correct answer. The data is in English, in text format.

What you can do with it

  • SFT models on math reasoning
  • Train a small reasoning model using R1 traces
  • Evaluate answer verification and reasoning quality
Languages
English
Size
100K–1M rows
Downloads this month
151.6K
Likes
823

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("open-r1/OpenR1-Math-220k")