HongGongLi/SpecFormer

该数据集是SpecFormer全光谱光学多层薄膜数据集的HuggingFace友好Parquet导出版本。它包含10,000,000个训练样本、500,000个验证样本和500,000个测试样本。每个样本存储了紧凑的uint16光谱目标，以及填充的结构标签和制造相关约束，适用于逆设计和光谱条件建模任务。774维目标按VIS-R、VIS-T、NIR-R、NIR-T和法向入射红外发射率排序。光谱值以uint16格式存储，通过1000缩放；除以1000可恢复[0, 1]范围内的归一化浮点光谱。This dataset is the Hugging Face friendly parquet export of the SpecFormer full-spectrum optical multilayer thin-film dataset. It contains 10,000,000 training samples, 500,000 validation samples, and 500,000 test samples. Each sample stores compact uint16 spectral targets together with padded structural labels and fabrication-related constraints for inverse design and spectrum-conditioned modeling tasks. The 774-dimensional target is ordered as VIS-R, VIS-T, NIR-R, NIR-T, and normal-incidence IR emissivity. Spectral values are stored as uint16 values scaled by 1000; divide by 1000 to recover normalized floating-point spectra in [0, 1].