CrystalReasoner/CrystalReasoner-MP

CrystalReasoner-MP数据集包含基于Materials Project的参考表，用于晶体结构生成、属性条件提示和评估。数据集以Parquet格式导出，包含200290行数据，列包括来自Materials Project的材料ID、结构、能量信息、稳定性、带隙、金属性、费米能、泊松比、描述和化学式，以及来自MLIP的体积模量、剪切模量和300K热膨胀系数。结构列以简单晶体文本格式序列化，部分属性列可能包含JSON序列化值或缺失值。数据集需下载并转换为本地shelve数据库以供训练和评估使用。CrystalReasoner-MP dataset contains the Materials Project-derived reference table used for crystal-structure generation, property-conditioned prompts, and evaluation. The dataset is stored as a Parquet export with 200290 rows, including columns from Materials Project such as material_id, structure, energy_above_hull, formation_energy_per_atom, is_stable, band_gap, is_metal, efermi, homogeneous_poisson, description, formula_pretty, and columns from MLIP like bulk_modulus, shear_modulus, thermal_expansion_300k. The structure column is serialized in a simple-crystal text format, and some property columns may contain JSON-serialized values or missing values. It is designed to be downloaded and reconstructed locally for training and evaluation purposes.