How Kumo Tabular Works

Kumo Tabular is a pre-trained structured-data foundation model. Instead of training a task-specific model, you give it labelled rows and the rows you want scored, and it answers in one inference pass.

One request type, for every model

sdm-client builds one Request type for both a single-table question and a multi-table one. A Request is the wire envelope minus the two things that belong to the call rather than the question: which model answers it, and which output fields you want back. Kumo Tabular is simply the case where the schema has one table and no relationships — not a second system.

labelled rows + rows to score → Request → SDMClient.predict(request, model=KumoTabular()) → PredictionResult

In-context prediction

A Request has two halves:

• context — rows where the answer is already known. The model learns from these.
• predict — rows being asked about, including rows the model has never seen.

The rows being scored need not appear in the labelled half. That is the point: Kumo Tabular exists to score rows it has never seen, without a training step.

The task kind is stated, never guessed

Kumo Tabular’s NIM accepts only the coarse classification or regression task kind and resolves binary versus multiclass itself from the context. sdm-client sends classification as given and never narrows it — inferring binary-versus-multiclass from a column’s cardinality is a guess, and it is wrong exactly when a binary target looks categorical.

What Kumo Tabular can and cannot read

Standard workflow