Overview

CSV is a lightweight training format for datasets with one input, a reference context, and one or more outputs.

Use the single-output shape for one output, or the multi-output shape when each row has multiple independent output labels.

Download templates

Download single-output CSV template Download multi-output CSV template

CSV shapes

Single output

The single-output shape contains:

• Input
• Output
• Reference

Column order does not change the required names. This shape is converted to a StateMachineInstruction.

For example:

Input	Output	Reference
Hello	greeting	A greeting
What should I wear?	advice	A request for advice

Multiple outputs

A multi-output CSV must contain:

• Input
• Reference
• At least one additional output column, such as OutputOne, OutputTwo, or Intent

For example:

Input	OutputOne	OutputTwo	Reference
Hello	greeting	positive	A greeting
What should I wear?	advice	neutral	A request for advice

The exact Input,Output,Reference shape is always treated as single output. A CSV with Input, Reference, and additional output columns is treated as multi-output and converted to a MultiClassifierInstruction.

Output-column order is preserved. Each output column gets its own state map containing its unique non-empty values.

Row rules

• The first data row must contain a non-empty value in every output column. Later empty output cells inherit the previous non-empty value; in multi-output CSVs, inheritance is tracked independently for each output column.
• Every data row must provide an output value directly or inherit it from the previous row.
• A multi-output row produces exactly one value for every declared output key.
• Input and Reference are required.
• Headers must be unique.
• At least one output column must exist in addition to Input and Reference.

The converter removes rows with an empty Input before processing them.

Multi-output

Multi-output samples use a JSON object keyed by output column name:

{"OutputOne":"greeting","OutputTwo":"positive"}

This allows each output column to select from the labels discovered in that column.

Errors

Conversion fails when Input or Reference is missing, headers are duplicated, no output column is declared, or the first data row has an empty output cell.

Where to go next

• Review the Columns section for column-by-column details.
• Use the Examples page to store full example CSVs.