Start a Data Science project

Use Data Science when notebooks explore reusable Python code. The preview creates a package, tests, and an output-free starter notebook. Jupyter is required; the scientific runtime stack is optional.

Start with notebooks and a package

uvx --from "create-forge[engine]==0.3.2" create-forge new "Notebook Study" --engine-preview --archetype data-science --capability jupyter --yes --data license=mit
cd notebook-study
uv run --locked poe check
uv run poe notebook

The final command starts JupyterLab. Open notebooks/getting-started.ipynb. Put reusable functions in src/notebook_study/ and tests in tests/, then import the package from your notebooks. Stop the server with Ctrl+C when finished.

Include the scientific stack

For numerical work, data frames, charts, or machine learning, generate a second project with Scientific Python selected:

uvx --from "create-forge[engine]==0.3.2" create-forge new "Model Study" --engine-preview --archetype data-science --capability jupyter --capability scientific-python --yes --data license=mit
cd model-study
uv run --locked poe check
uv run poe notebook

This also installs NumPy, pandas, Matplotlib, and scikit-learn as runtime dependencies and includes a test that imports them. No dataset, model, training pipeline, or deployment platform is generated.

Working files

Location	Intended contents
src/<package>/	Reusable Python code
tests/	Tests for package behaviour
notebooks/	Exploratory and reporting notebooks
data/raw/	Source data as received
data/interim/ and data/processed/	Intermediate and prepared datasets
models/	Trained or serialised models
artifacts/	Generated figures, reports, and outputs

The data, model, and artifact directories are ignored by Git and are not created during generation. Create them when needed. Keep credentials, datasets, and generated binaries out of version control.

Check notebooks before committing

Clear notebook outputs and execution counts, then run:

uv run --locked poe notebook:check
uv run --locked poe check

Notebook validation checks stored notebook content before executing temporary copies with the project's kernel. It leaves source notebooks unchanged. Execution is not sandboxed: only run notebooks you trust, since they have your user's filesystem and network access.

poe check already includes the notebook check, so use the dedicated task when you only need to validate notebooks. Both Data Science compositions are preview projects and cannot use create-forge update.

---

Missing an example or found an unclear step? Send documentation feedback.