DeepSeek-Prover-V1.5
https://github.com/deepseek-ai/deepseek-prover-v1.5- Category
- AI Tools
- Rank
- No. 1152Tools index
Previous survey · No. 1159 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- deepseek-ai
- GitHub
- 584 stars
- Date
About
DeepSeek's theorem-proving language model that combines RL and Monte Carlo tree search to solve formal math problems in Lean.
What it does
It proposes Lean 4 proof completions, checks each candidate with the Lean verifier, and can explore alternate proof steps when candidates fail. Three public model variants support raw generation, supervised fine-tuning, and reinforcement learning. The repository also supplies experiment orchestration, concurrent verification, result logging, and benchmark summaries.
Why it's ranked here
The strongest case is measurable proof-search performance plus reproducible machinery. The reinforcement-trained model with RMaxTS reaches 63.5% on miniF2F and 25.3% on ProofNet, exceeding every comparison shown. Still, those results come from demanding experiment settings and only two formal-mathematics benchmarks, so they do not establish broad mathematical coverage.
What's good
Lean feedback closes the loop between plausible text and accepted formal proof. The verifier rejects errors, incomplete goals, and uses of sorry. Search can preserve valid proof segments, expose tactic states, and pursue diverse continuations. Public base, supervised, and reinforcement-trained models also make controlled comparisons practical.
Tradeoffs
The supported setup is narrow: Linux, Python 3.10, Lean 4, a built Mathlib environment, pinned dependencies, and at least one CUDA-visible GPU for experiments. The published RMaxTS configuration uses many search processes, high verifier concurrency, large batches, and thousands of samples. That makes headline reproduction materially heavier than the quick example.
How to use it well
Use it for Lean 4 research, benchmark experiments, or workflows where generated proofs must pass machine verification. Start with direct generation and verification, then add sampling or RMaxTS when one-pass completion is insufficient. It does not replace Lean installation, Mathlib, model-serving infrastructure, or the work of formalizing an informal theorem statement.
Technical notes+
quick_start.py loads deepseek-ai/DeepSeek-Prover-V1.5-RL through Transformers and vLLM, generates one completion, extracts the Lean block, and verifies it through Lean4ServerScheduler. prover/lean/verifier.py invokes lake exe repl, separates errors, warnings, and sorries, and marks completeness only when no unresolved proof gaps remain. prover/launch.py requires a detected CUDA GPU and coordinates generator, verifier, and search processes. configs/RMaxTS.py selects RMaxTS with 6,400 samples, 256 search processes, 64 verifier workers, and tactic-state comments. prover/lean/proof.py segments accepted proof prefixes using parsed tactic positions and states. prover/summarize.py aggregates success artifacts by dataset split.
Observed
- Code license
- MIT License
- Model terms
- DeepSeekMath models use a separate Model License that permits commercial use
- Primary implementation language
- Python, with Lean 4 used for proof verification
- Install surface
- Clone with submodules, install pinned packages from requirements.txt, then build Mathlib4 with Lake
- Interfaces
- Python quick-start plus module-based command-line runners for search and result summarization
- Platform support
- Linux with Python 3.10; experiment launching requires at least one detected CUDA GPU
Read from README.md, requirements.txt, quick_start.py, prover/utils.py, prover/launch.py, configs/RMaxTS.py, configs/sampling.py, prover/summarize.py, configs/sampling_few_shot.py, prover/lean/proof.py, prover/lean/verifier.py, prover/workers/search.py, prover/algorithms/base.py, prover/lean/ast_parser.py, prover/workers/__init__.py.
What it can do
Prove mathematical theorems
Mathematical theorem statement in Lean → Formal proof in Lean syntax
Verify proof correctness
Formal proof in Lean → Verification result and error feedback
Generate proof steps using Monte Carlo tree search
Theorem statement and partial proof → Next proof step suggestions
Solve formal mathematics problems
Formal math problem in Lean → Complete solution with proof
Search proof space using reinforcement learning
Theorem goal and context → Optimized proof search path
Convert informal math statements to formal Lean code
Natural language mathematical statement → Formal Lean theorem definition
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.