CrystalReasoner/Qwen2.5-3B-CrysReas-NoEnergyTerm
CrystalReasoner/Qwen2.5-3B-CrysReas-NoEnergyTerm is a 3 billion parameter Qwen2.5-based language model developed by CrystalReasoner, specifically 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 excels at translating material descriptions into detailed lattice vector information and atomic coordinates, making it suitable for materials science applications.
Loading preview...
CrystalReasoner/Qwen2.5-3B-CrysReas-NoEnergyTerm Overview
This model, developed by CrystalReasoner, is an end-to-end LLM framework built upon the Qwen2.5-3B architecture, specifically designed for generating crystal structures. It processes natural language instructions to output detailed descriptions of crystal lattices and atomic positions.
Key Capabilities
- Crystal Structure Generation: Translates natural language descriptions of bulk materials (e.g., chemical formula, bulk modulus) into crystallographic data.
- Incorporates Physical Priors: Uses "thinking traces" during supervised fine-tuning (SFT) to integrate crystallographic and physical knowledge before generating coordinates.
- Reinforcement Learning (RL): Employs RL with verifiable rewards to improve the validity, stability, and property conditioning of the generated crystal structures.
- Output Format: Generates descriptions of lattice vector lengths and angles, followed by element types and coordinates for each atom within the lattice.
Good For
- Materials Science Research: Automating the generation of hypothetical crystal structures based on desired properties or chemical compositions.
- Computational Chemistry: Providing initial structural inputs for simulations or further analysis.
- Educational Tools: Demonstrating the relationship between material properties and atomic arrangements.