DVC Cheat Sheet

1Install & initset up

pip install dvc # or dvc[s3], dvc[gs], dvc[ssh]★
Install with the extras for your remote (dvc[s3], [gs], [azure], [ssh]…). Also via brew/conda/choco. Python 3.9+.
git init && dvc init★
DVC lives inside a Git repo. dvc init creates .dvc/; commit it. Run from the repo root.
git add .dvc .dvcignore && git commit -m "init dvc"
Commit the DVC config. .dvcignore works like .gitignore for DVC's scanning.
dvc version · dvc doctor
dvc doctor reports version, platform, and which remote backends are available.

2Track data (dvc add)the core idea

dvc add data/train.csv★
DVC moves the file to its cache and writes a small data/train.csv.dvc pointer (a hash). The real file is now git-ignored.
git add data/train.csv.dvc data/.gitignore git commit -m "track dataset"★
Commit the .dvc pointer + the auto-generated .gitignore to Git. Git now versions the pointer; DVC versions the bytes.
dvc add data/ # a whole directory
Track directories too — one .dvc file summarizes the whole tree.
dvc status · dvc checkout · dvc commit
status shows data drift vs. the pointer; checkout restores cached data into the workspace; commit updates a pointer after changing tracked data.

3Remotes & push/pullstore the bytes

dvc remote add -d storage s3://bucket/path★
Configure a default (-d) remote — where the actual data lives. Backends: S3, GCS, Azure, SSH, HDFS, HTTP, local dir.
dvc push★
Upload cached data to the remote (like git push for bytes). Do this after dvc add/repro + the Git commit.
dvc pull★
Download the data referenced by the current commit's pointers. The typical clone flow: git clone → dvc pull.
dvc remote modify storage endpointurl https://... # S3-compatible
Per-remote settings (endpoint, credentials, region). Use --local to keep secrets out of Git-tracked config.

4Access & import datareuse datasets

dvc get https://github.com/org/repo data/model.pkl
Download a DVC-tracked file from any DVC repo — no tracking, just fetch (like wget for DVC data).
dvc import https://github.com/org/repo data/train.csv★
Like get, but records the source so dvc update can pull newer versions later. Creates a .dvc import pointer.
dvc import-url s3://bucket/raw.csv data/raw.csv
Track an external URL/cloud path as a DVC dependency, re-fetchable with dvc update.
import dvc.api with dvc.api.open("data/train.csv", repo="...") as f: ...★
The Python API: stream a tracked file from any repo/version without downloading the whole thing. Also dvc.api.read(), get_url().

5Versioning & switchingtime travel

git checkout <commit> && dvc checkout★
The core workflow: Git switches the pointers, dvc checkout syncs the workspace data to match. Together = the exact state of that commit.
git tag v1.0 · dvc push
Tag data+code snapshots in Git; the matching data is in the remote. Reproducible releases.
dvc data status · dvc list <repo> <path>
Inspect tracked data changes; browse the DVC-tracked files of a remote repo.
# forgot dvc checkout after git checkout? data won't match the codegotcha
Git only moves the pointers — always follow a branch/commit switch with dvc checkout (or install DVC's Git hooks to automate it).

6Define stages (dvc.yaml)the DAG

dvc stage add -n train \ -d src/train.py -d data/train.csv \ -p lr,epochs -o model.pkl -M metrics.json \ python src/train.py★
Define a stage: -d deps, -o outputs, -p params, -M metrics, then the command. Written to dvc.yaml.
# dvc.yaml (edit by hand too) stages: train: cmd: python src/train.py deps: [src/train.py, data/train.csv] outs: [model.pkl]★
dvc.yaml is the pipeline definition. Stages chain into a DAG: one stage's outs are another's deps.
params: [lr, epochs] # read from params.yaml
Hyperparameters live in params.yaml; stages declare which keys they depend on so changes trigger reruns.
dvc dag · dvc stage list
Visualize the pipeline DAG in the terminal; list defined stages.

7Run & reproduceonly what changed

dvc repro★
Run the pipeline, re-executing only stages whose deps/params/outs changed. Results are cached — unchanged stages are skipped.
dvc repro train · dvc repro -f · dvc repro -s train
Reproduce up to a target stage; -f forces a rerun; -s runs a single stage without upstream.
dvc.lock # auto-written: exact hashes of deps/outs★
dvc.lock records the exact state of each run. Commit it with dvc.yaml so runs are reproducible.
dvc repro --pull · dvc repro --dry
--pull fetches missing deps from the remote first; --dry shows what would run.

8Params, metrics & plotsmeasure

dvc params diff · dvc metrics diff · dvc metrics show★
Compare params/metrics between commits or experiments. Metrics are small JSON/YAML files declared with -M/metrics:.
dvc plots show logs.csv · dvc plots diff★
Render plots (loss curves, confusion matrices) from CSV/JSON to an HTML report; diff overlays multiple revisions.
# dvc.yaml metrics: [metrics.json] plots: [logs.csv]
Declare metrics & plots in the pipeline so DVC knows what to compare across runs.

9Experiments (dvc exp)iterate fast

dvc exp run --set-param lr=0.01★
Run the pipeline as an experiment, overriding params on the fly. Experiments are lightweight — no manual commits/branches.
dvc exp show★
A table of all experiments with their params & metrics for side-by-side comparison. --sort-by/--drop to focus.
dvc exp apply <exp> · dvc exp branch <exp> <name>
Promote a good experiment into your workspace, or turn it into a real Git branch to keep.
dvc exp run --queue --set-param lr=0.1,0.01 dvc queue start -j 2
Queue a grid of experiments and run them (in parallel). dvc exp remove / dvc exp gc to clean up.
dvc exp push origin · dvc exp pull origin
Share experiments through your Git remote + DVC remote for collaboration.

10DVCLivelog from code

pip install dvclive from dvclive import Live★
DVCLive is the Python logger for metrics/params/plots — the code-side companion to dvc exp.
with Live() as live: for epoch in range(n): live.log_metric("acc", acc) live.next_step()★
Log per-step metrics; DVCLive writes files DVC reads for exp show/plots. live.log_param(), log_image(), log_artifact().
# auto-logging: PyTorch Lightning, Keras, HF, XGBoost callbacks
Framework integrations log automatically — drop in the DVCLive callback and metrics stream to DVC.

11Git integration & workflowtwo tools, one repo

# commit code + pointers to Git, data to the DVC remote git add . && git commit -m "..." dvc push★
The golden rule: Git for pointers/code, DVC remote for bytes. Always pair a Git commit with a dvc push.
dvc install
Installs Git hooks so git checkout auto-runs dvc checkout, and git push reminds you to dvc push.
# clone flow git clone <repo> && cd repo && dvc pull★
A collaborator gets code from Git and data from the DVC remote in two commands.
dvc gc -w # garbage-collect unused cachecare
Reclaim cache space. Scope it (-w = workspace) carefully — unscoped GC can drop data other branches need.

12Ecosystem & toolsthe wider stack

# VS Code extension: experiment table, plots, DAG in the IDE★
The official DVC VS Code extension gives a GUI for experiments, metrics, and plots.
# DVCLive + Studio: hosted experiment tracking & model registry
DVC Studio (Iterative) is the web UI for tracking experiments and a model registry on top of your Git repo.
# CML: CI/CD for ML (GitHub Actions / GitLab)
CML runs DVC pipelines in CI and posts metrics/plots as PR comments — from the same Iterative ecosystem.
# pairs with any training stack; remotes on all major clouds
Framework-agnostic and cloud-agnostic — a common backbone for reproducible ML/MLOps.

DVC cheat sheet

Outline

Data versioning

Sharing data

Pipelines

Experiments

Workflow & ecosystem

Data Versioning

Pipelines

Experiments

Workflow & Ecosystem

Worth memorizing

Outline

Data versioning

Sharing data

Pipelines

Experiments

Workflow & ecosystem

Data Versioning

Sharing Data

Pipelines

Experiments

Workflow & Ecosystem

Worth memorizing