Key Concepts: Datasets, Models, Labels

This article explains the core building blocks of Labelf. Understanding these concepts will help you get the most out of the platform.

A dataset is a collection of your customer interactions. Think of it as a large table: each row is an interaction (or a message within a conversation), and each column holds a specific piece of information about that interaction.

Datasets in Labelf have two types of columns:

Dataset columns are the fields that come from your source system. These vary depending on your data, but common examples include agent name, interaction type, resolution category, response time, language, and any custom fields from your support platform.

System columns are generated by Labelf when your data is imported. These include the utterance (the text content of the interaction), ContactId (a unique identifier linking messages to the same conversation), language detection, and cluster assignments.

Your workspace may contain one or more datasets. You'll select which dataset to work with when using Search or creating a new model.

A model is an AI classifier that Labelf trains on your data. Once trained and deployed, a model automatically categorizes new interactions into the labels you've defined.

For example, you might create a model called "Issue Resolution" that classifies every interaction as "Resolved", "Partially Resolved", or "Not Resolved". Or a "Customer Issue" model that identifies whether a conversation is about billing, technical support, account changes, or something else.

Labelf's classification models are powered by a large language model that analyzes entire conversations, considering the full context of a dialogue to make its prediction. You describe the classification task in plain language, define your labels with descriptions, and the model learns to categorize conversations based on those instructions — no pre-labeled data required.

Models go through several stages:

Models can be organized into folders on the Models page. A common approach is to group related models together — for example, putting all resolution-related models in a "Resolution" folder and all product-related models in a "Product" folder.

Labels are the categories that a model uses to classify interactions. When you create a model, you define your labels by giving each one a name and a description. You can also use AI to generate label suggestions based on your task description.

For example, an "Issue Resolution" model might have labels like "Resolved", "Partially Resolved", and "Not Resolved" — each with a description explaining exactly what qualifies.

Good labels are:

A playlist is a saved set of search results in the Search section. When you find a useful search query and filter combination, you can save it as a playlist so you can return to it later without recreating the search from scratch.

Playlists are helpful for ongoing monitoring — for example, saving a playlist for "all unresolved billing disputes from the last 30 days" that you check regularly.

A dashboard is a visual report made up of charts. Dashboards pull data from your datasets and models to show trends, volumes, distributions, and other metrics in an easy-to-read format.

Dashboards can be published (visible to your team) or kept unpublished while you're building them. They're the primary way to share insights from your data with stakeholders.

A workspace is the top-level container for everything in Labelf — your datasets, models, dashboards, and team members all live within a workspace. If your organization uses Labelf for multiple projects or departments, you might have separate workspaces for each.

You can switch between workspaces using the workspace menu in the top-right corner of the screen.

Here's how these concepts connect in a typical workflow:

If you have questions about any of these concepts, reach out to us at support@labelf.ai or use the chat widget in the bottom-right corner of the screen.