Datasets and runtime data
elfes.data brings physical samples into the tensor structures used by models. A dataset supplies individual samples, and a DataLoader combines them into batches while preserving atom order, basis references, and the structure of each physical quantity. Models receive the same data organization whether the selected values are read from disk or held in memory.
Datasets and loading describes how to read ELFES HDF5 files or convert physical samples already in memory. Samples and batches explains the shared geometry and basis fields and how indices change during batching. Physical data structures details how scalar, orbital, and real-space quantities fit into those samples.
Connectivity adds the model's directed edges when they are needed. Orbital data preparation selects requested matrix blocks and computes overlaps for a batch. Dataset statistics collects quantities such as neighbor counts and normalization moments over a chosen set of samples. These operations are explicit steps built on the loaded data; their physical meaning follows the definitions in elfes.physics.