Package a Reproducible Model

Reproduce the artifact

Record source commit, data version or immutable snapshot, feature code, label definition, split rule, random seeds, runtime, package lockfile, hardware assumptions, parameters, metrics, and artifact checksum. A notebook with hidden state is evidence of exploration, not a release package.

Run training from a clean environment. Confirm that the same command produces compatible metrics and an artifact that passes the same tests. Small numeric differences may be acceptable, but tolerances must be declared.

Define the inference contract

Specify input schema, units, allowed ranges, missing values, output schema, model version, timeout, error behavior, batch limits, and privacy rules. Validate outside the model. Add golden examples, boundary cases, malformed requests, and deterministic preprocessing tests.

Package preprocessing and postprocessing with the model or version them as explicit dependencies. Training-serving skew often lives outside the model file.

Produce a model card

Document intended use, prohibited use, training data context, evaluation sets, subgroup results, limitations, known failure modes, security and privacy considerations, owner, and rollback target. Link claims to versioned reports.

Practice: clean-room reproduction

Train a small public-data model from a fresh checkout. Register parameters, metrics, environment, and artifact. Start a local service, send valid and invalid requests, and verify the response schema. Ask another person or CI job to reproduce it from the written command.

Release boundary

Promotion should reference an immutable artifact and evaluation report. Never retrain silently during deployment or approve “latest” without knowing what changed.