training-pipeline

Input pipelines are part of the model system, not an afterthought. I keep dataset classes deterministic, move expensive transforms into explicit stages, and use DataLoader settings that match hardware limits. Good batching and collation logic can remo