reproducible-training

name	reproducible-training
description	Build reproducible training: a versioned prep->train->evaluate->register pipeline (not a notebook), experiment tracking (params/metrics/code/data/env), a model registry as source of truth, and leakage-free time-aware validation.

Reproducible Training

Pipeline, not a script

prep -> train -> evaluate -> register, with versioned inputs. A notebook isn't reproducible.

Track everything

Params, metrics, code version, data version, environment. An untracked experiment is an indefensible result.

Registry

Versioned models + metrics + lineage + model card = source of truth for promotion.

Validate honestly

No target leakage, no future info, time-aware splits for temporal data, test set used once. A leaked metric is a production letdown with a delay. Significance -> applied-statistics.