Workbench

Upload, train, review, take it away

A marimo notebook in the browser — reactive cells, UI elements like tables and buttons, and an AI assistant you give your own key to. The first cell installs pyborch; the rest is one review job: upload images → train on this tab's GPU → a review table with the labels the model doubts first → torch.onnx.export. Nothing is installed.

Open the workbench → The Jupyter notebook instead →

How it differs from the notebook

• Reactive. Change a cell and only what depends on it reruns. Pick a row in the review table and the cell below follows.
• UI elements. A file upload, a table, a button become Python values — no widget code.
• AI assistant. Put your own API key in the settings and cells can be generated and errors fixed on request. Code and errors leave; images do not.
• The same wheel. import borch_webgpu as torch boots borch.ts here too.

Worth knowing

• The kernel takes about ten seconds to come up (Pyodide and marimo download). Cells do not run by themselves: press the run button at the bottom right.
• A label is the part of the file name before the first _. Upload nothing and a synthetic three-class set runs instead. The two calls doing the work are the library's: torch.decode_images turns the files into an NCHW batch with labels from the names, and torch.suspects scores each label by how many of its five nearest neighbours in the model's features disagree with it — the review table's order.
• It lives inside the tab's memory and Pyodide's package range — the same list as the notebook page.

Measured by tests/browser/marimo_probe.py, nightly: opening the page, pressing run, and the moment each of the four sections (adapter, trained, review table, export) appears.