v0.6

CrystalFormer v0.6 is a unified autoregressive transformer foundation model for inorganic crystalline materials that supports both de novo generation (DNG) and crystal structure prediction (CSP) within a single probabilistic framework. By leveraging space group symmetry to compactly represent crystals, the model seamlessly switches between unconditional generation and formula-conditioned prediction — no architectural changes required.

v0.4.2

• add reinforcement fine-tuning methods to fine-tune the CrystalFormer using machine learning force field or property prediction model, more details can be seen at our paper
• add pmap for multi-gpu training
• fix some bugs

v0.3

update path in tests