DVC (Data Version Control)

Version datasets and ML pipelines with DVC. Tracks large data files in Git, defines reproducible pipeline stages, manages remote storage on S3/GCS/Azure, and compares experiments — Git for data and models.

This skill makes Claude a DVC specialist. It puts large datasets and model artifacts under version control alongside code, wires up dvc.yaml pipeline stages with dependencies and outputs, configures remote storage, and reproduces runs deterministically. Covers experiment tracking, metrics and plots, data registries, and CI integration so ML work stays reproducible.

When to use

Use when versioning large datasets or models, building reproducible ML pipelines with DVC, configuring remote data storage, or comparing experiment runs.

Examples

Track a dataset

Version large files

Set up DVC to track my 5GB training dataset with an S3 remote and add it to Git

Build a pipeline

Reproducible stages

Create a dvc.yaml with prepare, train, and evaluate stages so dvc repro reruns only what changed

Compare experiments

Track metrics

Configure DVC experiments to log accuracy and loss metrics and diff two runs