Note Unified collection of 2.1M deduplicated math problems (from 27.9M originals). Combines all 8 datasets below with 92.36% deduplication.

Note Largest contributor (67.6%). From XenArcAI/MathX-5M with 94.6% deduplication.

Note Second largest (14.9%). Diverse math reasoning from PRIME-RL/Eurus-2-RL-Data.

Note Third largest (10.5%). Includes solve_rate and domain metadata for curriculum learning.

Note Includes model difficulty ratings. From Skywork/Skywork-OR1-RL-Data.

Note Highest quality (0.44% duplicates). Includes problem_type metadata.

Note Medium-scale dataset. Originally 72K with 33.5% duplicates removed.

Note Competition math (AIME, AMC, Omni-MATH). High preservation rate (94%).

Note Smallest but fully preserved. From BytedTsinghua-SIA/DAPO-Math-17k.