Machine learning workflows

Work across the lifecycle

The command catalog separates ML data, experimentation, evaluation, and deployment. There are also LLM and retrieval workflows.

Choose a machine learning engineer profile for the task, then select the specific workflow. Profiles guide what to inspect and verify; commands define how to approach the job.

Plan compute explicitly

/just-vibe:ml-train Plan checkpointed training on one GPU for at most
six hours. Preserve the sample order when resuming. Do not provision compute.

Supply the model, data version, hardware constraints, budget, checkpoint policy, and objective when known. The agent should recover configuration from the repository where possible. Missing compute authorization is a limit on launching a run, not a reason to invent a budget.

Compare the evidence, not just the score

/just-vibe:ml-evaluate Compare these MLflow exports, including weak
segments. Check that splits and evaluation protocols are comparable.

Experiment helpers import supplied local exports. They can expose missing lineage, slice differences, and incompatible comparisons. They cannot establish causality from a collection of run metrics.

Keep the important questions visible

• Could labels, preprocessing, or temporal splits leak information?
• Are baselines and candidate models evaluated under the same conditions?
• Do aggregate scores hide weak or poorly represented segments?
• Can another run reproduce the relevant data and configuration?
• What happens to latency, drift, rollback, and cost after deployment?

These are workflow priorities, not a guarantee that a model is ready for production. Open a command page for its specific technical method, scope, and stopping conditions.