Custom training
Custom training is not how hosted pull-request reviews run today.
Hosted reviews call the model a platform administrator configured for the project. They run the official review/code package. They do not select a customer adapter. Managed inference of a trained adapter on hosted reviews is not available.
What you learn from review replies is review feedback. That memory is separate from this page.
What the Training page does now
Training (/team/<namespace>/training) lets owners and admins prepare a private catalog and save source filters:
- Link the private GitHub repository that holds adversaries and their history.
- Choose source repositories. The catalog repository is excluded.
- Optionally limit history with Merged on or after, Merged on or before, pull request authors, and reviewers. Usernames are comma-separated. Empty filters mean everyone and all dates.
The page states that collection is not available yet. Saving sources and filters does not start training, does not change the catalog, and does not change hosted reviews. Automatic collection and model recommendations are not running.
You can open Training before hosted inference is enabled. It is not a substitute for Repos.
Local doomer train
doomer train is a CLI workflow for grading a home-built package against pull-request history and drafting package edits. Official catalog packages are jury-only. It does not call the hosted review queue, and there is no CLI command that starts the managed training setup above.
That command is documented with the CLI. Use it only when you are authoring a package on the secondary path.