> ## Documentation Index
> Fetch the complete documentation index at: https://docs.b-bot.space/llms.txt
> Use this file to discover all available pages before exploring further.

# Training

> Create conversation examples and fine-tune a model for an Agent

# Training

Training turns example conversations into a fine-tuning job for an Agent. Each training session contains ordered user and assistant turns and keeps the job status with the session.

<Warning>
  Training is different from Knowledge. [Knowledge](/essentials/knowledge)
  retrieves information from uploaded documents when an Agent needs it. Training
  sends example conversations to a model provider to create a fine-tuned model.
</Warning>

## Before you start

Select the Agent you want to train. A new training session is linked to the selected Agent, and the session can use models configured for that Agent.

## Create or open a session

1. Open **Training**.
2. Select **New session**.
3. Enter a clear session name. You can also add a description.
4. Select **Create** to open the session.

The Training page lists existing sessions with their names, descriptions, number of examples, and current status. Select a session card to open it. Use the delete action on a card to remove a session you no longer need.

<Tip>
  Use one session for one behavior or response style. A focused name such as
  `Support escalation replies` is easier to review than a general name such as
  `Training 1`.
</Tip>

## Choose a model

In the session, choose a **Base model** from the models configured for the selected Agent. If no model is available, configure a model for the Agent before starting fine-tuning.

The chosen model is the starting point for the fine-tuning job. The Hub does not use Knowledge documents as training examples automatically.

## Add conversation examples

Each example is an ordered conversation made of user and assistant turns.

```text theme={null}
User: My order arrived damaged. What should I do?
Assistant: I am sorry your order arrived damaged. Please send the order number and a photo of the package so I can help with the next step.
```

Select **Add example** to create an example with one user turn and one assistant turn. Then:

* Enter the message text in each turn.
* Select the user or assistant role beside a turn.
* Select **Add turn** to continue the conversation.
* Delete a turn when it is not needed.
* Duplicate an example to create a similar variation.
* Delete an example to remove it from the session.

Write the response you want the Agent to learn. Keep roles correct, include enough context in the user turn, and avoid contradictory assistant responses.

## Save and start fine-tuning

Select **Save** to store the session name, description, and examples without starting a job.

When the examples are ready:

1. Confirm that the intended Agent is selected.
2. Confirm the base model.
3. Review every user and assistant turn.
4. Select **Start training**.

The Hub saves the current session before it starts the fine-tuning job. You can start only when at least one turn contains text and no job is currently active.

## Follow the job status

The session displays the current fine-tuning stage:

* **Draft**: The session has not started a job.
* **Queued**: The provider accepted the job and queued it.
* **Validating files**: The provider is validating the prepared training data.
* **Running**: Fine-tuning is in progress.
* **Deployment in progress**: The resulting model is being prepared.
* **Succeeded**: Fine-tuning completed. The resulting model identifier is shown.
* **Failed**: The job stopped with an error.
* **Cancelled**: The active job was cancelled.

An active job can be cancelled from its status panel. When available, the panel also shows the number of trained tokens or the provider error.

## Training and Knowledge

Use the feature that matches your goal:

* Use [Knowledge](/essentials/knowledge) for facts that should remain editable and retrievable from documents.
* Use Training for repeated examples of how an Agent should respond.
* Update a Knowledge document when source information changes.
* Create and review a new training session when you need another fine-tuned behavior.

## Next steps

<Card title="Knowledge" icon="book-open" href="/essentials/knowledge">
  Add documents that an Agent can retrieve as grounded context.
</Card>