Classification models are the core of Labelf's analytics engine. A model analyzes entire conversations in your dataset and categorizes them automatically — turning unstructured interactions into structured, actionable data. This article walks you through the model creation process.
How classification models work
Labelf's classification models are powered by a large language model that considers the full context of a conversation 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. This means you can go from idea to working model in minutes — no pre-labeled data required.
For even higher accuracy, you can optionally refine the model through prompt tuning or fine-tuning on annotated data (see [Improving Your Model]).
Before you start
Make sure your dataset has been connected and is showing as ready in Labelf. You'll also want to have a clear idea of what you're trying to classify — for example, issue type, resolution status, churn risk, or sales opportunity.
Starting the wizard
From the top navigation bar, click Models to open the Models page. Then click the + Create Classification Model button in the top-right corner.
This launches the model creation wizard, which guides you through the process step by step.
Step 1: Select Dataset
Choose which dataset your model will be built on. Each dataset shows its name, the number of items it contains, and its status. Select the dataset you want to use and click Next.
If your workspace has multiple datasets, make sure you pick the one that contains the interactions relevant to the classification task you have in mind.
Step 2: Apply Filters (Optional)
Filters let you narrow down which records from your dataset the model will work with. This is useful when you only want to classify a subset of your data — for example, interactions from a specific channel, time period, or language.
You can add two types of filters:
Column Filter — Filter by any column in your dataset. Choose between Include (only include records matching these values) and Exclude (remove records matching these values). For example, you could include only interactions where the language is English.
Model Filter — Filter based on the predictions of another deployed model. This lets you chain models together — for example, first classifying by topic and then building a resolution model only for billing-related interactions.
If you don't need to filter, simply click Next to continue with the full dataset.
Step 3: Select Label Column (Optional)
If your dataset already contains a column with label values that you want the model to learn from, you can select it here. The dropdown lists all available columns along with their data type (keyword, int, float, etc.). Once you select a column, Labelf shows a preview of the example values so you can verify it's the right one.
Most models are built from scratch — if you're defining your own labels in the next step (which is the typical workflow), you can skip this step and click Next to
continue.
Step 4: Configure Your Model
This is the most important step — where you describe your classification task and define your labels.
Task Description
Write a clear description of what the model should classify. Think of it as explaining the task to a new team member. For example:
"Determine the main reason the customer contacted support"
"Identify whether the agent attempted to sell a product to the customer"
"Assess whether the customer's issue was fully resolved by the end of the conversation"
The more specific your description, the better the model will perform.
Generate labels with AI
Click the Generate labels button to let AI suggest labels based on your task description. This is a helpful starting point — you can then customize, add, or remove labels as needed.
Define custom labels
Each label requires three things:
Label name — A short, descriptive name (e.g., "Billing Issue", "Technical Support", "Churn Risk")
Label description — A detailed explanation of what this label means. For example: "The customer is experiencing a problem with charges, invoices, or payment methods." Good descriptions are crucial — they directly guide how the model classifies conversations.
Color — Choose a color for the label, which will be used throughout the platform to visually identify this category.
Click + Add a new label to add more labels. You can remove labels by clicking the × button next to each one.
Step 5: Name Your Model
Give your model a name and optionally add more details:
Model Name (required) — Choose a descriptive name that makes it easy to identify what this model classifies. For example, "Issue Resolution", "Customer Intent", or "Churn Risk Detection".
Description (optional) — Add notes about what the model does or any context that will help your team understand its purpose.
Color (optional) — Pick a color to visually distinguish this model in the Models list.
Folder (optional) — Place the model in an existing folder, or leave it at the root level. Organizing models into folders helps keep things tidy as your workspace grows.
Click Next to continue.
Step 6: Review
The final step shows a summary of your configuration — dataset, labels, and model name. Review everything carefully. If anything looks wrong, click Back to return to earlier steps and make changes.
When you're ready, click Create Model.
What happens next
Once you click Create Model, you have a choice in how to proceed:
Deploy immediately — Your model can be deployed right away based on the task description and label definitions you provided. This gets you up and running quickly, though accuracy may improve with further tuning.
Prompt tuning first — Iterate on your task description and label definitions using a small validated dataset. This lets you refine the model's instructions before full deployment. See [Improving Your Model] for details.
Annotate and fine-tune — For maximum accuracy, you can annotate a training set and fine-tune the model on your labeled data. This additional step teaches the model the specific patterns in your conversations.
You can monitor your model's status on the Models page as it progresses through Training → Validating → Deploying → Ready/Deployed.
Next steps
Improving Your Model — Learn about prompt tuning and annotation to boost accuracy
Understanding Model Metrics — Interpret your model's performance metrics
Deploying Your Model — Learn about the deployment lifecycle and using predictions
Need help?
If you have questions about creating models, reach out to us at support@labelf.ai or use the chat widget in the bottom-right corner of the screen.