CrystalReasoner/Qwen2.5-3B-CrysReas-Base
CrystalReasoner/Qwen2.5-3B-CrysReas-Base is a 3.09 billion parameter language model developed by CrystalReasoner, fine-tuned for generating crystal structures from natural language instructions. This model utilizes supervised fine-tuning (SFT) and reinforcement learning (RL) to incorporate crystallographic and physical priors, enhancing the validity, stability, and property conditioning of generated structures. It specializes in translating material descriptions into detailed lattice vector and atomic coordinate data, making it suitable for materials science applications.
Loading preview...
Overview
CrystalReasoner/Qwen2.5-3B-CrysReas-Base is an end-to-end large language model (LLM) framework designed for the generation of crystal structures based on natural language input. Developed by CrystalReasoner, this model leverages a multi-stage training approach to ensure high-quality and physically plausible outputs.
Key Capabilities
- Natural Language to Crystal Structure Generation: Translates descriptive text about bulk materials (e.g., chemical formula, bulk modulus) into detailed crystallographic information, including lattice vectors and atomic coordinates.
- Enhanced Reasoning: Incorporates "thinking traces" during supervised fine-tuning (SFT) to embed crystallographic and physical priors, improving the structural validity before generating coordinates.
- Reinforcement Learning (RL): Utilizes RL with verifiable rewards to further refine generated structures, focusing on improving their validity, stability, and adherence to specified properties.
- Pymatgen Integration: Provides a script to convert the generated structural data into a
pymatgen Structureformat, facilitating downstream materials science analysis and simulations.
Good For
- Materials Scientists: Ideal for researchers and engineers needing to quickly generate candidate crystal structures based on desired properties or compositions.
- Computational Materials Design: Useful for accelerating the initial stages of materials discovery and design by automating the generation of structural inputs for simulations.
- Educational Purposes: Can serve as a tool for understanding the relationship between material properties and atomic arrangements.