Train adapters
An adapter is a small model trained for one task in your calls. This section explains what Mubit does with your recorded calls, from the first call of a task to an adapter that answers some of them, and what you can see and change at each step. It is for developers whose calls already go through Mubit.
How it works
Nothing here needs a code change. Mubit does each step on its own, and the console shows where each task is.
- Tasks. Mubit groups calls that share their instructions, tools, tool choice, output format and model into a task.
- Ratings. Mubit rates each task's new calls automatically, good or bad. You can rate a call yourself in the console or from code, and your rating replaces Mubit's.
- Training. When a task has 25 calls rated good, and enough of them to keep some back for an exam, Mubit trains an adapter version for it.
- The exam. Mubit asks the new version the conversations it kept back, and compares its replies with the replies your provider model gave.
- The rollout. A version first answers copies of calls in a silent test, then a share of the task's calls in a partial rollout, then the task's calls in full.
- Answering calls. For a call of the task, Mubit can answer with the adapter instead of your provider. Your app sends the same request with the same model ID, and gets a reply in the same format.
On app.mubit.ai today, a version that scores below its exam's target can still go into the rollout, and Mubit does not take a low-scoring version offline on its own. The exam and The rollout say what this means and what you can do.
In this section
What makes a task, which calls are kept, and how calls are rated.
When a task trains, on which base model, and when a new version trains.
What a version is tested on, its target, and each result.
Silent test, partial rollout and live. Rolling back and taking a version offline.
Which calls an adapter can answer, and how to tell who answered a call.
What runs on your provider key, and where call content goes.
The limits you can meet today.
A task that doesn't train, and a version that doesn't answer calls.
Start here
Send examples of one task and follow it to its first version.
How calls form a task, and how sessions group the calls of one conversation.
Bring in past calls from Langfuse, Braintrust or Pydantic Logfire.
The Concepts page explains each term. Training is on for your workspace only where the console shows an Adapters page. To ask about it, see Get help.