Marimo Example Notebook

marimo example notebooks for wandb/examples

Read ../marimo-notebook/SKILL.md first for the marimo file format, reactivity rules, and marimo check. If converting an existing Jupyter notebook, also read ../jupyter-to-marimo/SKILL.md. This skill layers repo-specific conventions on top of those.

The canonical exemplar is examples/marimo/mnist-registry/mnist_registry.py — when in doubt, match its structure.

Repo conventions

• Each example lives in its own directory: examples/marimo/<example-name>/, with the notebook as <example_name>.py.

• The .py file is the source of truth. A workflow generates the markdown export; never hand-edit a generated .md next to a notebook.

• Start the file with a PEP 723 script header (pinned lower bounds, e.g. "marimo>=0.9", "wandb>=0.18") followed by a module docstring that says what the notebook builds and how to run it:

  """One-paragraph summary of what the notebook builds.
  
  Run:
  
      uvx marimo edit <example_name>.py --sandbox
  """
  

• Runtime droppings (data/, wandb/, artifacts/, __marimo__/, model weights) must not be committed.

Notebook structure

Order the notebook as a narrative the reader scrolls through top to bottom:

1. Setup cell — with app.setup(hide_code=True): holds all imports, constants, and environment detection (e.g. device selection). One place, not scattered across cells.
2. Intro prose — a markdown cell (hide_code=True) with the title, "What you will build", and "Prerequisites". Prose goes in markdown cells, never mixed into code cells.
3. mo.outline() in its own cell near the top, so readers see the notebook's shape at a glance.
4. Configuration — all UI controls batched into a single submittable form (see below).
5. The pipeline — logic cells that consume the form, interleaved with ## Section markdown cells.
6. Verify and next steps — a closing markdown cell telling the reader exactly what to look at (in the W&B UI: which charts, tabs, panels) and what to try next.
7. Helper functions — @app.function / @app.class_definition cells under a ## Helper functions section at the bottom.

Gate execution once, then let the graph run

Batch every control into one form so nothing expensive runs until the user submits:

form = (
    mo.md(
        """
        **Training.**

        {epochs}  {batch_size}
        ...
        """
    )
    .batch(epochs=epochs, batch_size=batch_size, ...)
    .form(submit_button_label="Train model", bordered=False)
)
form

form.value is None until submit. Gate one cell on it, with a message that tells the reader what will happen:

mo.stop(
    form.value is None,
    mo.md("Fill in the form above and click **Train model** to ..."),
)
cfg = form.value

Every downstream cell references names defined after the gate (cfg, run, model, ...), so marimo's dependency graph holds them all back until the form is submitted. Do not re-check the button/form in later cells, wrap cells in if guards, or use mo.ui.run_button() when a form fits — one mo.stop() replaces all of that.

Separate logic from presentation

• Heavy lifting (loading data, training, logging, saving artifacts) goes in named @app.function helpers; the cell body becomes a short, readable call: model, history, final_acc, best_acc = run_training(...).
• View cells (hide_code=True) render results and contain no logic worth reading.
• Push temporaries into functions to keep notebook globals to a minimum — marimo notebooks work best with few globals, and every global name is reserved across the whole file.
• Present results with real components — mo.ui.table(rows, selection=None) for tabular results, mo.callout(..., kind="success"/"warn"/"danger") for status, mo.vstack for grouping — not markdown with emoji.

W&B integration patterns

• Auth: offer a mo.ui.text(kind="password") API-key field that falls back to ambient login (wandb login, WANDB_API_KEY, netrc) when blank. Never write the key into the run config.
• Re-runs: marimo keeps the kernel alive across form re-submits, so finish any prior run first: if wandb.run is not None: wandb.finish().
• Entity: include an entity field and explain that accounts created after May 2024 have no personal entity — the run must go to a team.
• Surface the run URL immediately after wandb.init so readers can watch metrics stream: mo.md(f"**Run started:** [{run.name}]({run.url})").
• Expected failures become guidance, not tracebacks. Wrap only the calls that fail for account-setup reasons (wandb.init, registry linking) and render a mo.callout(kind="danger") that names the likely cause and the fix. Let everything else fail naturally — no try/except for control flow.
• A recoverable step (e.g. registry linking) should capture its outcome in a status value that a separate view cell renders, so the pipeline completes either way.
• Group metrics into UI sections with slash-prefixed names (Training/loss), and put headline numbers in run.summary.

Before handing back

• uvx marimo check <notebook.py> passes.
• Globals audit: anything only used inside one step should live in a helper.
• The notebook reads top-to-bottom as a tutorial; every code cell either shows output or is a named helper.
• Fresh-eyes test: a reader with a new W&B account can follow Prerequisites, submit the form, and verify the result from "Verify and next steps" alone.