| | --- |
| | license: apache-2.0 |
| | datasets: |
| | - ZJUVAI/GenesisGeo |
| | base_model: |
| | - Qwen/Qwen3-0.6B-Base |
| | --- |
| | |
| | # GenesisGeo Model |
| |
|
| | [📃Paper](https://arxiv.org/abs/2509.21896) • [📚 Github](https://github.com/ZJUVAI/Newclid) • [📊 Dataset](https://huggingface.co/datasets/ZJUVAI/GenesisGeo) |
| |
|
| | This model is **specialized in automated geometric theorem proving**, capable of proposing auxiliary constructions to solve challenging geometry problems. It forms the core of the [**GenesisGeo**](https://github.com/ZJUVAI/Newclid) project—a neuro-symbolic system that reproduces the AlphaGeometry framework using the Newclid infrastructure. |
| |
|
| | Developed through large-scale synthetic training, this model demonstrates strong performance in geometric reasoning tasks. It is built upon the Qwen3-0.6B-Base architecture, fine-tuned specifically for generating auxiliary points and constructions in complex proof scenarios. |
| |
|
| | ## Model Description |
| |
|
| | - **Architecture:** Transformer-based language model |
| | - **Base Model:** Qwen3-0.6B-Base |
| | - **Training Dataset:** [GenesisGeo Dataset](https://huggingface.co/datasets/ZJUVAI/GenesisGeo) (21.8 million synthetically generated geometric theorems) |
| | - **Purpose:** Proposing auxiliary constructions in geometric proofs within a neuro-symbolic reasoning loop |
| |
|
| | ## Performance |
| |
|
| | The integrated neuro-symbolic system achieves: |
| | - **24/30** problems solved on the IMO-AG-30 benchmark |
| | - Close to the original AlphaGeometry performance (25/30) |
| |
|
| |
|
| |
|
| |  |
| |
|
| |
|
