Lean Workbook: 140,124 Formalized Math Contest Problems

140,124 contest-level math problems formalized in the Lean 4 theorem prover, created by internlm and released in October 2024. The dataset includes natural language statements, answers, formal statements, and formal proofs where available. It is intended to support the training of autoformalization models and automated proof search.

Use Cases

• Train autoformalization models based on the paired natural language and formal statements.
• Search for automated proofs based on the provided formal statements and proofs.
• Benchmark theorem provers on a large corpus of formalized contest problems.

Strengths

• Contains 140,124 total problems across two splits (57,231 and 82,893).
• Provides four key components per problem: natural language statement, answer, formal statement, and formal proof (if available).
• The dataset and associated code are open-source.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for individual splits is provided, but total file size and specific formats are unknown.
• Last updated 2024-10-09 07:15:23; freshness should be verified.

Provenance

Source

huggingface

Collection Method

Likely compiled from contest-level math problems and formalized by the authors.

Freshness

Last updated 2024-10-09 07:15:23.