Skip to content
voker-logo voker-logo Dashboard

Fine-Tuning

Fine-tuning creates a custom version of a model that learns directly from your own examples. Instead of relying solely on a foundational model like gpt-4o-mini that’s trained on broad data, fine-tuning helps your Voker respond in your product’s voice, follow complex instructions, and deliver more consistent, predictable outputs.

When to Use Fine-Tuning

a11y.sectionLink When to Use Fine-Tuning
  • Want domain-specific accuracy
    • For example, your Voker should understand your company’s terminology or industry-specific phrasing.
  • Keep fixing similar mistakes
    • If you’re repeatedly reprompting the Voker to fit your preferred tone, format, or structure, fine-tuning helps the model learn those patterns automatically.
  • Have strong example data
    • When you already have, or want to have, a collection of high-quality responses that show what “good” looks like, you can teach the model to follow them consistently.

If you’re still experimenting with prompts, formats, or overall behavior, we recommend sticking with foundational models, as it’s more flexible for iteration.

Using Fine-Tuning

a11y.sectionLink Using Fine-Tuning

1. Create a Dataset

a11y.sectionLink 1. Create a Dataset

Start by creating a new dataset in your workspace. This will store your examples.

Create a new empty dataset Create a new empty dataset
Enter the dataset name for fine-tuning Enter the dataset name for fine-tuning

2. Run Test Executes

a11y.sectionLink 2. Run Test Executes

Use Test Execute to experiment with your Voker. This playground lets you test your Voker’s behavior with your configured instructions, tools, and output schema.

Configure your Voker’s instructions to reflect the tone, style, and rules you want the model to learn.

Brand guideline Voker instructions Brand guideline Voker instructions

Run a Test Execute to generate outputs that demonstrate your brand’s style and behavior.

Running a test execute of the brand guideline Voker Running a test execute of the brand guideline Voker

3. Add to Dataset

a11y.sectionLink 3. Add to Dataset

When you see a result you like, click Add to dataset. Each saved example becomes one row in your dataset. You’ll need at least 10 rows before you can train a model.

Add examples directly from Test Execute into your dataset for training.

Add test execute outputs to your dataset Add test execute outputs to your dataset

4. Train a Model

a11y.sectionLink 4. Train a Model

Once your dataset reaches 10+ rows, click Train Model to start fine-tuning. The training process will create a new fine-tuned model that learns from your collected examples.

Train model success Train model success

5. Attach to Voker

a11y.sectionLink 5. Attach to Voker

After training completes, go to the Configure page and select your fine-tuned model under the “Fine-tuned models” tab.

Attach the fine-tuned model to a new Voker to apply your learned tone and structure.

Attaching the fine-tuned model to a new Voker Attaching the fine-tuned model to a new Voker

Example: Testing the Difference

a11y.sectionLink Example: Testing the Difference

Below is an example of a Policy Chatbot that uses the fine-tuned brand-guideline model. The fine-tuned version consistently applies Voker’s brand voice — using approved terms, polished phrasing, and tone alignment across responses.

With Fine-Tuned Model

Policy chatbot response using fine-tuned model Policy chatbot response using fine-tuned model

With Base Model

Policy chatbot response using base model Policy chatbot response using base model

The base model ignores brand-specific phrasing and lacks tone consistency.