workflow/

The tools that actually run machine learning.

One module per tool, each declaring its ToolSpec next to its executor. These are the calls an agent chains together to get from a dataset to a served model, and the only group whose results become notebook cells.

Tools

tuiml_train: Fit a model, with optional preprocessing, feature

selection and cross-validation. Returns a model_id.

tuiml_predict: Predict with a previously trained model.
tuiml_evaluate: Score a trained model on a dataset.
tuiml_benchmark: Compare several algorithms over the same data.
tuiml_tune: Grid or random hyperparameter search.
tuiml_plot: Confusion matrix, ROC, PR, learning curve, CD diagram.
tuiml_save_model: Write a trained model to a path of your choosing.

Notes

Most of these take a model_id produced by tuiml_train. Ids resolve through an index rehydrated from ~/.tuiml/models/ at import, so one survives an MCP server restart.