Train adapters

Troubleshooting adapters

This page goes from what you see in the console to what to do. It is for anyone whose task or adapter isn't doing what they expect. Start from the task's page in Tasks, or the adapter's page in Adapters: their status says most of it.

A task doesn't train

Point at the task's status on Tasks, or open the task: its page says why in words.

What you see	Why	What to do
`Collecting examples`, with a count such as `12 of 25 examples`	The task has too few kept examples.	Send more calls of the task, or rate more of its calls good.
`Waiting for more examples: {k} of {need}`	The task has 25, but too few to keep 12 back for the exam.	Send more calls of the task, with new conversations.
`Waiting: today's training limit`	The workspace's training budget for the day is spent.	Nothing. Training starts after 00:00 UTC.
`Waiting: the pause between trainings`	Mubit pauses between two trainings of a task.	Nothing. Training starts after the pause.
`Waiting for new examples`	A training failed.	Send more calls of the task. Mubit tries again later.
`Can't train`	The first training failed. The page says why after `No adapter yet:`.	See the next table.
No count goes up	The calls aren't recorded in full, or they don't form one task.	Check the calls in Requests, and see Tasks and sessions.

When the first training fails:

Reason	What to do
`too few examples to train on`	Send more calls with new messages.
`examples too long to train on`	Send calls with shorter requests and replies.
`the training run failed`	Send more calls with new messages. Mubit tries again a day later.
`the first training stopped`	Send more calls with new messages.

A version stays below target

Rate the task's calls: a reply rated bad is never trained on, and your ratings replace Mubit's. See Rate a call yourself.
Send more varied calls of the task, so the next version has more to learn from.
Keep the text that changes per call in the user message, not in the instructions, so the calls stay one task.

A version below target can still answer calls today. See How the exam acts today.

A version doesn't move on

What you see	Why
It stays `Not live yet`	Another version of the task is still in its rollout, or the task's calls can't go to an adapter. See Calls answered by an adapter.
It stays in `Silent test`	It has fewer than 50 compared replies from 10 conversations yet. A silent test gets copies of up to 100 calls a day.
It stays in `Partial rollout`	It has answered fewer than 100 calls in its partial rollout yet.

An adapter never answers calls

Check, in order:

The call's API and key. Only Chat Completions calls on /openai/v1, with an OpenAI key or a Neon AI Gateway key, can go to an adapter.
Recording. A call sent with x-middleware-record: metadata or off, or x-middleware-store: false, never goes to an adapter.
The version's step. A version in a silent test answers only copies. One in a partial rollout answers its share of conversations.
The adapter's switch. An adapter that reads Switched off or Stopped answers nothing until an admin switches it on, or makes a version live.
The call itself. A call with an image or a document, or one longer than the adapter's context, goes to your provider. So does a call Mubit isn't confident belongs to the task.
Availability. When the adapters' serving capacity isn't ready, your provider answers. This passes on its own.

Requests, filtered on Served by Your adapter, lists the calls an adapter did answer.

A version was taken offline

On app.mubit.ai today, Mubit doesn't take a version offline on its own. A version reads Taken offline after someone took it offline, or after an emergency stop. The adapter's history shows each move with its reason. An admin can move the version to a silent test again, or make another version live.

An adapter answers badly

Stop it from answering: an admin uses Roll back…, Take offline… or Switch off… on the adapter's page. See The rollout.
Rate its bad replies Bad in Requests, so the next version doesn't learn from them.
Compare its replies with your provider model's in Requests: turn on Include copies to see the spot-check copies.

Get help

If none of this explains it, see Get help. Quote the task's name and, for a call, its request ID.