How do you make an ML experiment reproducible?

Direct answer

What a strong answer covers

How it works

An experiment is reproducible when another person, on another machine, can produce the same artifact and metrics from the recorded inputs. The record includes the exact commit, the data as an immutable snapshot or a versioned query with its result hash, the feature definitions and their versions, every hyperparameter, the random seeds for each library, the container image digest with pinned dependencies, and any hardware assumption such as GPU nondeterminism. The proof is not the record but a rerun in a clean environment that is compared artifact hash to artifact hash and metric to metric, with tolerances stated where numerics legitimately vary.

Example

A researcher’s notebook result could not be reproduced by the platform team until the run was moved into a tracked pipeline: the data query was pinned to a table version, seeds were set for NumPy, PyTorch, and the data loader, the image digest was recorded, and the rerun matched the metric to four decimal places. A remaining 0.1 percent variation was traced to nondeterministic GPU kernels and documented as the tolerance.

Common mistakes

• Training from a query against a live table that changes daily.
• Recording hyperparameters but not the data version or the environment.
• Calling an experiment reproducible without ever rerunning it.

A useful answer structure

Answer from your own work.

Do not memorise a script: use only examples you can explain honestly in follow-up questions.