Skip to main content
A trained version is the output of a training run. A registered model is the long-lived operator-facing track you use to organise versions, decide what should stay active, and manage ongoing rollout decisions. The important mental model is:
  • Events define what the system should learn from
  • Templates package those signals into a repeatable training recipe
  • Training runs create new trained versions
  • Models are optional long-lived tracks you use to organise versions over time
  • Serving targets decide which approved version is live for a given tenant or surface

​
How To Think About Training

Users should think about training and retraining as a versioning workflow, not as editing one monolithic model in place.
1

Create or update signals

Create or update a signal template on the Events page.
2

Start training

Start a training run from that template.
3

Create a trained version

Let the platform create a new trained version.
4

Review the result

Review the result on the Models page.
5

Add a label

Optionally add a human-friendly label and description.
6

Approve the version

Approve the trained version when it is ready.
7

Promote to serving

Promote it to a serving target when you want it live.
8

Repeat when retraining

Repeat the same flow whenever you retrain.
That means retraining does not overwrite the previous result. It creates another version you can compare, approve, reject, or promote.

​
Training And Retraining Flow

​
System Name vs Human Label

Every trained version has two layers of identity:
  • A system name generated from the template name and the training run identifier
  • An optional label and description added by your team
Example:
  • System name: Engagement booster v2 · train-42-1712912345
  • Label: Homepage recommender
  • Description: Tighter freshness weighting for spring campaign
The system name preserves provenance. The label and description make the version easier for operators to recognise later.

​
Registered Models vs Trained Versions

Use these two concepts differently:
  • Trained versions are the raw outputs of training runs
  • Registered models are long-lived tracks you create when you want an explicit operator-owned model record for rollout, version management, or ongoing retraining policy
A common pattern is:
  • Train several versions from one template
  • Review and compare them
  • Promote one version
  • Later create a registered model if you want a named track for that recommendation strategy

​
What The Models Page Shows

The Models page brings together:
  • Successful trained versions from recent runs
  • Approval state
  • Serving targets
  • Which version is active
  • Registered models, if you use them as long-lived tracks
This lets you operate even if you are a brand-new account with no pre-existing model records.

​
Approve A Trained Version

After a training run completes successfully, the new trained version has a status of Pending. You must approve it before it can be promoted.
1

Open training or models

Open Console > Training Jobs or Console > Models.
2

Approve the completed version

Find the completed version and click Approve.
3

Confirm approved status

The version status updates to Approved.
Approving a version does not automatically change what is served. It makes that version eligible for promotion.

​
Promote A Version To Production

Once approved, you can promote a version to replace what is currently live on a serving target.
1

Find the approved version

Find the approved trained version.
2

Start promotion

Click Promote to production.
3

Choose a serving target

Choose the serving target you want to update.
4

Watch target status

The console will show the target status updating.
5

Confirm production traffic

Once the target returns to a healthy state, the promoted version is live and subsequent recommendation requests use it.
You can promote a previous version at any time if you need to roll back.

​
Active Version

The Models page shows which version is currently active on each serving target. That gives you a clear answer to:
  • Which version is live?
  • Where is it live?
  • When was it last promoted?
If no version is active yet, the target remains unassigned until you promote one.

​
Training run metrics

Each training run records a set of metrics at completion. Open a run from the Training Jobs page to view:
  • Training job status and duration
  • Final metrics (loss, accuracy, or custom metrics your training job reports)
  • The run manifest, which records the configuration used
Use these metrics to compare versions and decide whether a new one is ready to approve or promote.

​
Trigger A New Training Run

See the Configuring Events and Signal Templates guide for instructions on defining signals, saving templates, and creating the next trained version.