Benchmark and predict.

Benchmark

With one labelled file, the run holds out a stratified 20% by the target and scores it against the remaining 80% as context. Because the held-out rows have labels, it returns quality metrics:

• Accuracy, ROC-AUC and PR-AUC, with the ROC and precision–recall curves.
• F1 and a confusion matrix.
• The positive rate of the target and its imbalance, which drive AUC, PR-AUC and the score.

The held-out scores download as a CSV — the row id, the score and the true outcome — to set against your own model's scores on the same rows; History keeps them.

Predict

With a training file and a test file, the run scores every test row against the training rows as context. Test rows have no labels, so there are no quality metrics: you get a score per row, the most likely class and its confidence, the predicted class distribution, and the rows the model was least sure of. The predictions download as a CSV.

Benchmark, then predict

Uploading both files runs both: the benchmark first, to estimate how reliable the scores are, then the test file.

No training run

zypl.Foundation learns the labelled book in context. Nothing is trained for the run; the labelled rows are the context the test rows are scored against.

Time and cost

Every run records the context and test sizes, the compute time, the device it ran on, and its cost, with and without the worker's cold start. The demo scores on a hosted GPU worker. Each account can start a limited number of runs a day; the button that starts a run says how many are left.