> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://humanloop.com/docs/v4/guides/overview/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://humanloop.com/_mcp/server.

> Datasets are pre-defined collections of input-output pairs that you can use within Humanloop to define fixed examples for your projects.

Datasets are pre-defined collections of input-output pairs that you can use within Humanloop to define fixed examples for your projects.

A datapoint consists of three things:

* **Inputs**: a collection of prompt variable values which are interpolated into the prompt template of your model config at generation time (i.e. they replace the `{{ variables }}` you define in the prompt template.
* **Messages**: for chat models, as well as the prompt template, you may have a history of prior chat messages from the same conversation forming part of the input to the next generation. Datapoints can have these messages included as part of the input.
* **Target**: data representing the expected or intended output of the model. In the simplest case, this can simply be a string representing the exact output you hope the model produces for the example represented by the datapoint. In more complex cases, you can define an arbitrary JSON object for `target` with whatever fields are necessary to help you specify the intended behaviour. You can then use our [evaluations](/docs/v4/guides/evaluation/overview) feature to run the necessary code to compare the actual generated output with your `target` data to determine whether the result was as expected.

![Datapoints are pre-defined input-output pairs.](/docs/_fern-img/f2ca637640a0e5f66d6fe63f079201bfa25e27835ba41adf45f8bdb8516734f4.webp)

Datasets can be created via CSV upload, converting from existing Logs in your project, or by API requests.

> Datasets are collections of datapoints which represent input-output pairs for an LLM call.