MATH-500: 12,000 Training and 500 Test Math Problems for Step-by-Step Verification

12,000 training problems and 500 test problems from the MATH benchmark created by OpenAI for their 'Let's Verify Step by Step' paper. The dataset is hosted on HuggingFace by user sxiong and includes a training split not present in the original repository. The dataset card was last updated on April 11, 2026.

Use Cases

• Training language models for mathematical problem-solving based on the described benchmark problems.
• Benchmarking model performance on step-by-step verification tasks based on the MATH benchmark.
• Fine-tuning models for educational AI applications based on the structured math problems.

Strengths

• Contains 12,000 training problems, providing a substantial base for model training.
• Includes 500 test problems for standardized evaluation.
• Derived from the established MATH benchmark created by OpenAI.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the training split is specified, but total dataset size and file formats are unknown.
• License information is unknown, which may restrict usage.

Provenance

Source

OpenAI's MATH benchmark from the 'Let's Verify Step by Step' paper.

Collection Method

Likely extracted from the source repository mentioned in the description.

Time Range

null

Freshness

Last updated 2026-04-11 18:30:14; freshness should be verified.

Geography

null