How deepseek-ai/deepseek-prover-v2 works
DeepSeek-Prover-V2 is a released Lean 4 theorem-proving model, accompanied by a benchmark and MiniF2F proof archive. The README documents both a recursive proof-search process used to create training data and a user workflow that loads a released model with Hugging Face Transformers to generate proofs. The repository tree contains no implementation source, so the architecture below reflects documented responsibilities and relationships only; the README is the evidence for those links.
Main entry points: README.md Prover-V2 model · minif2f-solutions.zip MiniF2F solutions
Components 14
Proof Data Creation
- Theorem problem
README.md - DeepSeek-V3
README.md - Subgoal decomposition
README.md - 7B subgoal search
README.md - Proof synthesis
README.md - Cold-start data
README.md
Model Training
- Fine-tuning
README.md - Reinforcement learning
README.md
Inference and Evaluation
- Prover-V2 model
README.md - Hugging Face Transformers
README.md - Lean 4
README.md - ProverBench
README.md - MiniF2F solutions
minif2f-solutions.zip
Other components
- Model user
Connections 14
- Theorem problem → DeepSeek-V3: prompts
README.md - DeepSeek-V3 → Subgoal decomposition: decomposes
README.md - DeepSeek-V3 → 7B subgoal search: formalizes steps
README.md - Subgoal decomposition → 7B subgoal search: defines subgoals
README.md - 7B subgoal search → Proof synthesis: resolves subgoals
README.md - Proof synthesis → Cold-start data: pairs proofs and reasoning
README.md - Cold-start data → Fine-tuning: fine-tunes
README.md - Fine-tuning → Reinforcement learning: continues with
README.md - Reinforcement learning → Prover-V2 model: produces
README.md - Model user → Hugging Face Transformers: loads model
README.md - Hugging Face Transformers → Prover-V2 model: runs inference
README.md - Prover-V2 model → Lean 4: generates Lean proofs
README.md - Prover-V2 model → MiniF2F solutions: proofs available as
README.md - ProverBench → Theorem problem: formalizes problems
README.md