Fine-tuning your AI

The fine-tuning room is a dedicated space for you to try out your AI and test its responses. You'll find it on the menu under 'Knowledge'.

What is the purpose of fine-tuning?

Fine-tuning lets you test how your AI responds and add new prompts and completions straight to your training data. It's useful for:

• Testing specific questions, statements, or scenarios on different topics to see how your AI replies.
• Quickly adding new prompts and completions to your AI's training data.
• Experiencing what it's like to interact with your AI as a client.

Before you start

Fine-tuning is for testing and polishing, so your AI needs the right foundations in place first. Before you begin, make sure you have completed the style & tone and engagement rooms, so your AI knows how to communicate and guide conversations, and that you have given your AI any essential knowledge to draw on.

If you’re not sure what’s left to complete, head to your dashboard for suggestions.

Test your AI

Hold a chat with your AI as if you were a client, and see how it responds. To get the most from testing:

• Focus on one topic area before moving on to another.
• Have conversations that mirror how your clients might interact with your AI.
• Start a fresh conversation regularly by clicking the circular arrow icon, especially after adding or editing training data. Changes apply to new messages straight away, but your current chat can briefly still reflect the old data.
• Refresh the chat if you want to select a different starter prompt.

Add responses to your training data

If your AI gives a response you'd like to add to your training data, either as it is or with some changes, take these steps:

• Hover over the response and click the edit icon.

• In the 'Add to training data' window, edit the completion as needed.

• Also check that the prompt makes sense on its own, out of the context of the conversation, and edit it if needed. For example: if the prompt is one word such as "yes", you might change it to something more specific such as: "Yes, I'd like help preparing for my next training session", so the saved response is used in the right situations.

• Click 'Add'.

The new prompt and completion pair will be added to your training data, and you'll see it in both your fine-tuning history and training data overview table.

View, edit, and export your fine-tuning history

To view your fine-tuning history, toggle this on at the bottom of the page. You'll then see a table of the prompt and completion pairs you have added in fine-tuning.

These also appear in your training data overview table, with the source 'fine-tuning'. The two are linked, so editing or deleting a prompt or completion pair in either place updates the other.

To edit any prompt or completion here, click in the text box, make your edits, and click anywhere outside of the text box to save the changes.

To export your fine-tuning history, select all or individual rows, then click the export button.

Check your fine-tuning history regularly for conflicting or out-of-date responses, and remove any that no longer apply.

AI response issues

If your AI is not responding in fine-tuning as you'd like, follow these steps:

1. Check that you have saved any training data additions or edits before testing.
2. Refresh the chat in fine-tuning (by clicking the circular arrow) to ensure it's using the up-to-date data.
3. Review your fine-tuning history and check for out-of-date responses.
4. Search your training data for similar phrases to check for gaps or conflicting information.

Head back to earlier training rooms if you notice:

• Consistent style issues > Adjust your tone of voice description and style sliders in style & tone.
• Problems with greetings or initial responses > Set or update your conversation phrases in style & tone under 'advanced personalization'.
• Incorrect answers to starter prompts > Ensure you have added your ideal answers for each starter prompt in the engagement room.
• Responses that don’t follow your coaching approach > Add or update your coaching methodology in the engagement room under ‘Advanced coach mode’.
• Missing knowledge > Upload more content in knowledge upload, or add prompts and completions directly to training data.

Testing and improving further

Once you've tested your AI and made any refinements, it's a good idea to ask some trusted clients, contacts, or colleagues to try it out too. Not only can you gather valuable feedback from your testers, but you'll also be able to review their conversations to help spot gaps in your AI's training data.

Check out our guide on running an alpha test.