| license: cc-by-nc-4.0 | |
| library_name: pytorch | |
| tags: | |
| - crystal-structure-prediction | |
| - flow-matching | |
| - molecular-crystals | |
| # PackFlow checkpoints | |
| Curated PackFlow model weights from | |
| [PackFlow: Generative Molecular Crystal Structure Prediction via Reinforcement Learning Alignment](https://arxiv.org/abs/2602.20140). | |
| Load them with the [`packflow`](https://github.com/learningmatter-mit/packflow) package (`packflow download` / `load_checkpoint`). | |
| | File | Role | | |
| |------|------| | |
| | `packflow-2M/best_model.pt` | ~2M base model | | |
| | `packflow-20M/best_model.pt` | ~20M base model | | |
| | `packflow-ddp/best_model.pt` | ~60M canonical base model (`packflow-60M`) | | |
| | `packflow-pa/best_model.pt` | GRPO preference-aligned finetune of the 60M model | | |
| Training structures are **not** included. CSD refcodes for the splits are in the GitHub repository (`data/refcodes/`). | |
| ## License | |
| The **source code** for PackFlow is released under an MIT License. However, since PackFlow was trained on data from [CCDC's Cambridge Structural Database](https://www.ccdc.cam.ac.uk/), the **model weights** are released under a [Creative Commons Attribution-NonCommercial 4.0 International](https://creativecommons.org/licenses/by-nc/4.0/) (CC BY-NC 4.0) License. For commercial use of the model weights, please ensure that you have a proper [CCDC License](https://www.ccdc.cam.ac.uk/support-and-resources/licensing-information/). | |