OpenBMB released MathForm-8B, an autoformalization model for Lean 4
OpenBMB released MathForm-8B, an autoformalization model that translates natural-language math into Lean 4, on Hugging Face under Apache-2.0. It was trained on the FormalVerse dataset via supervised fine-tuning and reinforcement learning with Lean compilation and semantic-consistency feedback. The release includes a paper, GitHub repository, and evaluation pipeline.
China context
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the model weights from Hugging Face and integrate MathForm-8B into Lean 4 workflows using the provided Transformers, vLLM, or SGLang deployment options.
- For investors
- The Apache-2.0 license and open-weights release may attract commercial interest in autoformalization tools, but no funding or market traction data is provided in the evidence.
MathForm-8B is an autoformalization model that translates natural-language mathematical statements into Lean 4. It is released with the paper 'MathForm: Scaling Mathematical Autoformalization with Knowledge Retrieval and Verification-Guided Refinement'. The model is trained on FormalVerse through supervised fine-tuning followed by reinforcement learning using Lean compilation and semantic-consistency feedback. The release includes usage instructions for Transformers, vLLM, and SGLang, with recommended settings of temperature 0.6, top p 0.95, and max new tokens 16384. Evaluation scripts and benchmark files are available in the MathForm repository, and compilation checks require a running Kimina Lean Server with Lean 4.21.0. The project is licensed under Apache License 2.0.
MathForm-8B uses a pipeline that combines Mathlib knowledge retrieval, compilation and semantic verification, and iterative refinement to generate formal data, followed by trajectory reconstruction and training. The model is evaluated using Pass@8 pass rates under Syntax Check (SC) and Consistency Check (CC) across six benchmarks, with results shown in the model card. The model supports deployment via Transformers, vLLM, and SGLang, with an OpenAI-compatible API endpoint.
Researchers and developers working on formal mathematics gain an open-weights model that automates translation of natural-language math into Lean 4, potentially reducing manual formalization effort. The Apache-2.0 license allows commercial use, which could lower barriers for integrating autoformalization into proof assistants and mathematical software.
MathForm-8B provides a permissively licensed tool for autoformalization, which may be valuable for organizations developing formal verification, mathematical software, or AI-assisted theorem proving. The open-weights release enables integration into existing Lean 4 workflows without licensing costs.
The next verifiable signal is whether the MathForm repository's evaluation pipeline is independently run and whether the model's Pass@8 results are reproduced by third parties. Adoption can be tracked through GitHub stars, forks, and downstream usage of the FormalVerse dataset.