Examples

This page provides example CSV files to help you understand the format and structure.

Download Full Examples

Download CSV Chat Model Example Download Databiomes Chat Model Example

CSV Template Basic Structure

Input	Output	Reference
Explain this file to me	EXPLAIN	This is a template CSV file used for training models.
Explain this file in detail	EXPLAIN_DETAIL	This is a template CSV file used for training models. The file has three columns: Input, Output, and Reference. Reference is not used in the output, but is used to synthesize similar inputs for model training purposes.

CSV Chat Example

The following table shows example rows from the CSV chat example file:

Input	Output	Reference
What are the rules for input?	INPUT_RULES	The Input column requires a header and each row must provide its own input text. Input values are not inherited from previous rows. The maximum length for an Input cell is 300 characters. Each input should be a concrete, natural-language question or instruction.
What are the input column requirements?	(blank - inherits from row above)	(blank - optional)
What are the different columns in the CSV?	CSV_COLUMNS	Every CSV uses three required columns: Input, Output, and Reference. All three columns must be present in the header row. The Input column contains user-facing prompts, Output contains the model's intended response, and Reference contains human-written training context.
What columns does the CSV need?	(blank - inherits from row above)	(blank - optional)
What are the rules for output?	OUTPUT_RULES	The Output column requires a header and the first row must have a non-blank Output value. Cell values can be blank to inherit the most recent non-blank Output above. The maximum length for an Output cell is 300 characters. Output represents the model's intended response.
How does the output column work?	(blank - inherits from row above)	(blank - optional)
What are the rules for reference?	REFERENCE_RULES	The Reference column requires a header and cell values can be blank. A minimum of 10 total Reference rows are required. The maximum length is 300 characters. Reference is only used for training context and is not returned by the model's output. It helps the model understand context and synthesize similar inputs.
What is the reference column used for?	(blank - inherits from row above)	(blank - optional)
How does output inheritance work?	OUTPUT_INHERITANCE	When an Output cell is blank, it automatically inherits the most recent non-blank Output value from the row above. This allows multiple inputs to share the same output without repeating the value. The first row must always have a non-blank Output value to establish the initial output.
Can output values be inherited from previous rows?	(blank - inherits from row above)	(blank - optional)

Guardrail Example

These rows are from the Databiomes chat CSV example and show how to mark guardrails. The Output must be GUARDRAIL, and the following rows inherit the output.

Input	Output	Reference
What is the capital of Spain?	GUARDRAIL	(blank)
Who will win the next american election?	(blank - inherits from row above)	(blank)
where can I buy a nice couch?	(blank - inherits from row above)	(blank)

Download the full file to see all examples.