OpenBMB introduces mathform framework to advance automated mathematical formalization

The OpenBMB team has unveiled a new open-source framework, dataset, and model designed to enhance the automated formalization of mathematics. The project, named MathForm, aims to address the challenge of enabling machines to accurately interpret and verify mathematical theorems through formalization, a key hurdle in achieving general artificial intelligence. The framework introduces a retrieval-augmented and verification-guided data construction mechanism, which first extracts necessary definitions from the Mathlib library and then refines the output through multiple iterations based on feedback from the Lean compiler and semantic consistency checks. To support this initiative, OpenBMB has also launched the FormalVerse dataset, containing over 367,000 verified Lean4 examples across various mathematical domains. Experimental results indicate that models trained on FormalVerse achieved a 60.32% consistency check rate, significantly outperforming FineLeanCorpus at 46.53% and NuminaMath-LEAN at 41.49%. The MathForm-8B model, a core component of the project, demonstrated strong performance in six benchmark tests, achieving an 88.06% syntax check pass rate and a 72.37% consistency check pass rate on an 8-point scale. Notably, the MathForm-8B model outperformed larger models such as ReForm-32B and Goedel-Formalizer-V2-32B, despite having only a quarter of their parameter count. In the most challenging subsets, FATE-H and FATE-X, the model achieved consistency check success rates of 63% and 37%, respectively, surpassing the strongest specialized benchmarks by 10 and 12 percentage points. This highlights the model's strong reasoning and formal error correction capabilities. The project is now available on GitHub at https://github.com/openbmb/MathForm.