DeepMath-103K: 103,000 Verifiable Math Problems for RL and SFT

DeepMath-103K contains 103,000 math problems designed for reinforcement learning and supervised fine-tuning, released by zwhe99 in April 2025. The collection focuses on challenging, verifiable reasoning tasks and is associated with ArXiv paper 2504.11456. It was specifically updated in May 2025 to address answer leakage in 48 samples.

Use Cases

• Supervised fine-tuning (SFT) of language models using the math problem and solution pairs
• Reinforcement learning (RL) training utilizing the verifiable answers for reward signal generation
• Benchmarking mathematical reasoning capabilities across 103,000 distinct samples

Strengths

• 103,000 records
• MIT licensed for open use
• Decontaminated to remove answer hints in problem statements
• Verifiable solutions suitable for RL reward modeling

Limitations

• Narrowly focused on mathematical reasoning, limiting utility for general-purpose language tasks
• Lack of explicit column documentation in the primary metadata

Provenance

Source

zwhe99, ArXiv 2504.11456

Time Range

2025

Freshness

Updated as recently as May 29, 2025, to fix answer leakage issues.