NuminaMath-LEAN is a large-scale dataset of 100,000 mathematical competition problems formalized in the Lean 4 theorem prover language. It was created by AI-MO and is derived from a challenging subset of the NuminaMath 1.5 dataset, focusing on problems from competitions like the IMO and USAMO. The dataset was last updated on July 31, 2025.
Use Cases
- Training automated theorem provers based on the formal statements and proofs.
- Evaluating the performance of formal reasoning models on competition-level problems.
- Fine-tuning large language models for mathematical reasoning using formalized Lean 4 code.
- Studying the structure of high-difficulty mathematical proofs from competitions like the IMO.
Strengths
- Contains 100,000 formalized problems, described as the largest collection of its kind.
- Focuses on a challenging subset of problems from prestigious competitions.
- Data is human-annotated for formal statements and proofs.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is known, but specific file formats and data structure details are unknown.
- Data may reflect bias inherent to its source, focusing on specific competition problems.
Provenance
- Source
- AI-MO, derived from the NuminaMath 1.5 dataset.
- Collection Method
- Problems are formalized in the Lean 4 theorem prover language.
- Time Range
- null
- Freshness
- Last updated 2025-07-31 15:18:45; freshness should be verified.
- Geography
- null