DVC ML Workflow

DVC turns a git repo into a full ML lab: data and model files are versioned out-of-band (in a cache + remote), pipelines are declared in dvc.yaml, and experiments are run as ephemeral git commits with metrics and plots attached. No tracking server, no separate database — everything lives in your existing git history.

This skill is opinionated about the parts of DVC that matter for production ML work: pipelines, queued experiments, metrics/commit binding, and remotes. It defers to the official docs at https://dvc.org/doc for everything else and links them inline so the agent always pulls the latest guidance.

When to use

• User wants reproducible ML pipelines without a tracking server (mlflow, wandb, etc.)
• User mentions dvc.yaml, params.yaml, dvc exp run, dvc queue, dvc push, .dvc/cache
• User wants to do a hyperparameter sweep / grid search and have each run land as a separate commit with metrics
• User wants to version a dataset or model file too large for git
• User asks "how do I make my training reproducible" and is already on git
• User wants mlflow ui-style experiment comparison but doesn't want to run a server (DVC's dvc exp show + VS Code extension fills that role)

When NOT to use

• User wants a hosted experiment dashboard with multi-user collaboration → use mlflow-tracking skill
• User wants LLM trace observability (spans, prompts, token costs) → DVC has no story here; use mlflow-tracking