Git for datasets, Makefiles for models, no server required

What it does

DVC is a command-line tool and VS Code extension that versions data and machine learning models without stuffing them into Git. It keeps lightweight metadata in your repository while storing the actual files in an external cache, and it tracks ML experiments locally using nothing but Git history. Pipelines define how models are built from code and data, skipping steps that have not changed.

The interesting bit

Rather than funneling you into a managed SaaS platform, DVC turns your existing Git repository and generic cloud storage into a full experiment-management system. The VS Code extension adds a GUI for browsing experiments and data, though the README admits pipeline support there is still on the roadmap.

Key highlights

• Data and model files live in an external cache; only version metadata enters Git, keeping the repository lean
• Pipelines act as computational graphs that connect code, data, and outputs, rerunning only what changed
• Experiment tracking requires no central server; hyperparameters, metrics, and plots are compared through local Git state
• Supports standard remote storage including S3, Azure, Google Cloud, and SSH-accessible on-premise servers
• A VS Code extension provides visual experiment tracking and data management

Caveats

• The VS Code extension currently lacks data pipeline support; the README notes it is “coming soon”
• The extension and the core tool are separate installs, so the GUI will not work out of the box without the CLI backend

Verdict

A solid fit for ML engineers who want reproducibility and data versioning without adopting another hosted dashboard. Probably overkill if your datasets are small enough for vanilla Git and you do not need incremental pipeline runs.