personalitylinmult.data

Training-data layer: the FI HDF5 artifact, its builder, and its loaders.

This package owns everything between the feature extractors in personalitylinmult.preprocess.fi and the training loop in personalitylinmult.train:

  • schema — the typed feature/dataset contract.

  • sources — readers for each on-disk format.

  • builder — the config-driven HDF5 writer.

  • dataset — random-access reads from that HDF5.

  • collate — batching with per-feature masks.

  • datamodule — the Lightning DataModule.

The HDF5 file is the single training artifact and the reproducibility boundary: training reads it and nothing else. It is published to a public Hugging Face dataset repo (make push-fi-h5) so a run reproduces without re-extraction, and rebuilt from source by make pipeline-fi.

Submodules