DeepSeek-Prover-V2 is an open-source large language model designed for formal theorem proving in Lean 4, with a sophisticated data collection and reinforcement learning process for enhancing formal mathematical reasoning.
Features
Recursive Theorem Proving
Utilizes DeepSeek-V3 to decompose complex mathematical problems into subgoals and formalizes these steps for improved proof generation.
Reinforcement Learning Integration
Implement a cold-start training using synthesized data to better bridge informal reasoning with formal proof construction.
Model Variants Available
Offers models in two sizes: 7B and 671B parameters for various computational needs.
ProverBench Dataset
Includes a benchmark dataset with 325 formalized mathematical problems for comprehensive evaluation.
Hugging Face Compatibility
Easily deploy and utilize the model using Hugging Face's Transformers library for inference.