Start a training run

You decide when your models train. Once you have made enough annotations or classifications, you start a training run by clicking the Start Training button. Training does not begin in the background on its own.

Where the button is

There is one Start Training button per trainable model.

Classifiers:

• Legacy Classifier: in the classifier's Model Training status pill, shown in the Recommendations area on the Build page.

• Helix Classifier: in the classifier's Model Training status pill, on the Split & Classify page.

  

Extractors: Each document type's annotation page, in the header bar (top-right area).

In addition, the Model Training status pill hosts a Start Training action. The pill appears next to each trainable model or document type across the application, which means you can start an extractor training without navigating to its annotation page.

The button is not shown on the Build homepage as a standalone control, but the status pill on each document-type card does include the action.

How to start a training

1. Navigate to the model you want to train. Options:
   • Open the Split & Classify page (Helix Classifier only).
   • Open the classifier's annotation flow (Legacy Classifier).
   • Open the annotation page for the document type (extractor).
   • Or, find the Model Training status pill for that model (for example, on the Build homepage or Measure overview) and use its Start Training action.
2. Check the changes counter next to the Start Training button or inside the pill. This shows how many annotations or classifications have accumulated since the last training.
3. Select Start Training. The status changes to Queued.
4. The system picks up the queued training and starts the run. This may take a few minutes.
5. When the run completes, the status changes to Trained and shows the updated score, last training date, duration, and base model version.

Button states

The button's enabled state depends on what has changed since the last training and whether a training is already in flight.

ステート	ボタン	Popover message
Enough changes	有効	No popover. Select to queue a training.
Below the minimum changes threshold	無効	"At least N changes are needed before a new training can be started."
No changes since the last training	無効	"No changes have been made since the last training."
Training queued or in progress	無効	"A training is already queued or in progress for this model."

What counts as a change

Each annotation or classification modification counts as one change. For example, annotating a field on a document or classifying a page both count. The changes counter resets after a successful training run. Document type schema or base model changes count as major changes and bypass the required change threshold.

Status pills

Status pills appear next to each trainable model or document type across the application. The pill reflects the current state of that model's training, and (where applicable) exposes the Start Training action.

ステート	What you see
Not yet trained	Changes counter. Start Training is enabled once the threshold is reached.
Queued	Message: "Training is being prepared and will start automatically. This may take a few minutes."
Training in progress	In-flight indicator. Start Training is not shown.
失敗	Error message, warning icon, and a Retry button to re-queue the training.
Trained	Last training date, duration, and base model version used. Changes counter since the last run. Start Training is enabled if enough new changes exist.

Recommendations and warnings

• When a large number of changes have accumulated, the status pill shows a warning icon to nudge you to start a new training.
• When you change the document type schema, the change is not included in the trained model until you start a new training. A warning is shown.
• When the base model version changes, the existing trained model is not re-aligned automatically. A warning is shown, and you start a new training when ready.

例外

• Zip import. When you import a zip into a project, a training is queued automatically. You do not need to select Start Training.
• One training at a time. You cannot queue a second training for the same model while one is already queued or running. Wait for it to finish (or fail) before starting another.