TrainingPreview
Try it on 1% of one job. If it stumbles, your usual model takes over.
Answer support questions
What it learns from
The cleaned data
Only an admin or the owner can look at and approve the data. Nothing is sent to train on before that.
Each run is tested on your examples it never saw. It can be switched on only if it does better there than the model it was trained from.
Give it 1% of one job, or any share up to all of it. Everything else stays as it is, and each request takes one path from start to end.
If it starts failing, your usual model answers and the trained one rests. Cadreen tries it again after about 5 minutes, then after longer waits up to an hour, and brings it back after 5 good answers in a row.
When Cadreen holds the trained model, an admin can download it, once they confirm the licence of the model it started from. Every download is recorded, and if a record can't be written, Cadreen is told.
What your workspace has done in Cadreen, a file of examples you upload, a table in your own database, or records from a connected service. A person approves the cleaned data before training starts, and the model is tested on examples it hasn't seen.
An admin picks one job and how much of it the model gets. Each switch is kept with who made it and when.
Switch it off with a reason, and the job goes straight back to what it used before.
Today you set it up in the dashboard, or call the HTTP API. SDK, CLI and MCP support is on its way.
Yes. Training is in preview: a workspace admin can start a run from Training in your dashboard.
A file of examples of work done well, as a .jsonl file with one example per line, and at least 400 of them usable. Or the work Cadreen has already recorded for you, once there are at least 10,000 finished pieces of it, most of them done well.
Training is free while charges are off.
What your workspace has done in Cadreen, a file of examples you upload, a table in your own database, or records from a connected service. An admin gives each database or service to one workspace, and only that workspace's runs can read it. Before outside data is used, an admin confirms you have the right to train on it.
Before you switch it on, ask for a side-by-side: your model and the one you use now answer the same 30 questions yours wasn't trained on, and a third model says whose answer is closer to yours. It's advice, not a test. The real test is a small share of your own requests.
Only a workspace admin or the owner, signed in. An API key can't. If Cadreen pauses a model, a person's choice still wins.
After 5 pauses in a day, Cadreen stops trying it and says so on the job's page. The model stays saved; switch it on again when you're ready.
Yes, when Cadreen holds it. An admin first confirms the licence of the model it started from, and every download is recorded (if a record can't be written, Cadreen is told).
Start with a small share of one job. Your usual model takes over whenever it stumbles.