Introduction

​

Overview

Datasets can be used as data collections. One way to add data to a dataset is by exporting spans from both production traces and evaluation traces.

To help you filter traces and export only relevant ones to datasets, you can add labels to spans. Laminar provides you with a UI for labeling. This page explains label types.

​

Label classes

Every label on a span has a class. The class defines the name of the label, and the possible values. For example sentiment can be a class with possible values positive, negative, and neutral.

Every user can assign one label value per label class per span.

​

Internal representation

Internally, each possible label value is represented by a number. This is helpful to be able to add label values alongside evaluation scores.

​

Creating a label class

When you create a label class, you will need to assign a number to each possible label value. Make sure to use unique numbers for each possible label value.

For example, if you have a label class sentiment with possible values positive, negative, and neutral, you can assign the numbers 1, 2, and 3 to each of them.

​

Label description

Label descriptions are optional and can be used to provide more context about the label. We recommend setting them for better alignment between labelers, both human and LLM-as-a-judge labelers.