#7272nd of 80 in Open modelsFor science, health and research, science, research and papers
alphageometry
Research code for DDAR and AlphaGeometry olympiad geometry theorem provers
- Stars
- 4.9k
- License
- Apache-2.0
- Last commit
- Jan 2026
- Language
- Python
4,783 on 22 Apr 20264,903 on 11 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Reproduces the two geometry provers from the 2024 Nature paper: DDAR, a symbolic deduction engine (DD plus algebraic reasoning), and AlphaGeometry, which adds a 150M-parameter transformer that proposes auxiliary constructions via beam search in JAX. It prints numbered, fully justified proofs and ships the IMO-AG-30 and JGEX-AG-231 problem sets. DDAR solves 14 of 30 IMO-AG-30 problems and AlphaGeometry solves 25.
Who it is for: Researchers reproducing neuro-symbolic geometry theorem proving results
Strengths
- DDAR runs standalone without meliad or sentencepiece, so no checkpoint is needed
- Proofs are printed step by step with numbered, traceable predicates
- Includes benchmark problem files and a test suite for checking reproduced results
- Checkpoint and vocabulary are provided via download.sh under CC BY 4.0
Weaknesses
- Paper-scale results need 4 V100 GPUs and 250 CPU workers; parallelism code was removed
- Meliad must be cloned manually because it is not on pip
- Issue #14 reports LM scores differ between users, though top beams match
- README says Google will not maintain or support the code
What it needs
- GPU optional
- Needs meliad, JAX, Flax, TensorFlow, gin-config, SentencePiece, Python 3.10
- Models: AlphaGeometry 150M transformer checkpoint
Also in Open models
See all 80| Rank | Project | Score |
|---|---|---|
| 1 | ruviewWiFi CSI sensing on ESP32 for presence, breathing, heart rate and pose | 94 out of 100 |
| 2 | layaSingle-pass multilingual classifier returning typed decisions with calibrated odds | 88 out of 100 |
| 3 | rf-detrReal-time transformer models for object detection, segmentation and keypoints | 81 out of 100 |
| #4 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 78 out of 100 |
| #5 | open_clipPyTorch library for training and running CLIP-style image-text models | 78 out of 100 |