Models

Overview

The Models page shows the live state of the AI evidence chain and every model version recorded for it. From here you can see which model each service runs per language, start and stop training, activate a trained version, and inspect or delete versions.

Open the page

You need the Modify Settings permission. Click Administration in the navigation bar, then click the Models card. If you are already in Administration, click Models in the section bar.

Page layout

Click to enlarge

• 1 Status line - a summary of the four services, for example “4 serving” or “1 serving · 2 training or busy · 1 offline”, and when they were last checked.
• 2 Auto refresh - how often the page re-checks the services.
• 3 Check now - reloads the service status and the version list.
• 4 Training settings - opens AI settings.
• 5 Service tile - one tile per AI service in the Evidence chain, in pipeline order.
• 6 Service state - the state in words with a coloured light, for example Serving.
• 7 Languages - whether the service runs the active model for each language. Show languages lists the models.
• 8 Start training - opens the training dialog.
• 9 Model versions - search, filters, and the list of versions, grouped into trained versions and baseline defaults, below the tiles.

Status line and refreshing

• Check now reloads the service status and the version list immediately.
• Auto refresh re-checks every 10, 30, or 60 seconds, or with Auto (30s while active) only while a service is busy. When it is set to Off (manual), the page still re-checks on its own every 30 seconds while any service is training or busy, so a running training stays current.
• If the services can’t be reached, a message explains it and offers Try again; training controls stay locked until the services answer.

Evidence chain

The four services appear in the order documents pass through them: Topic detection → Entity recognition → Coreference → Relation extraction. Each tile shows the service name used in the deployment (for example entity-recognition), and these sections:

Service state

The state is written out in words, with a coloured light:

Training run

While a service trains, its tile shows the language and version being trained, how long the run has been going, and a bar measured against the 24-hour limit. A run still marked as training after 24 hours is marked failed.

If a version is still marked as training in the database but the service isn’t running it, the tile says so. It is marked failed once the 24-hour limit passes, or you can delete it from the list.

Languages

A strip of five language cells (EN, RO, HU, RU, FR) shows whether the service runs the model that is active for each language:

• Check — the service runs the active model.
• Exclamation — the service runs a different model; the sentence below says how many languages differ.
• Dash — not checked right now (the service is offline or busy).

Click the strip or Show languages to see, per language, the active model and, when they differ, the model the service actually runs.

Start and stop training

Each tile has one action:

• Start training opens the training dialog (see below). It is unavailable, with the reason shown, while the service is offline or busy, or while a version is still marked as training.
• Stop training appears instead while the service trains. A confirmation, Stop … training?, names the run; click Stop training to confirm; the version is marked stopped and active models don’t change.

Training dialog

Click to enlarge

• 1 Language options - the language to train, each with the model that is active for it.
• 2 What happens - the version the run creates, that nothing is activated automatically, and the time limits.
• 3 Cancel - closes the dialog without training.
• 4 Start English training - starts the run; the label names the chosen language.

The dialog is titled Train followed by the stage name.

1. Pick the language. Each option shows the model that is currently active for it.
2. The dialog shows what will happen: the new version it creates (for example ner_en_v2), that nothing is activated automatically, that training may take several hours, and the 24-hour limit.
3. Click Start training. If the service refuses, the reason appears in the dialog and you can try again or cancel. Esc or Cancel closes it.

Model versions

Search and filters

Click to enlarge

• 1 Search - matches name, model ID, stage, language, and status.

• 2 Stage - the stage filter, shown open.

• 3 Language, Sort, and Rows - the other filters, the sort order, and the page size.

• 4 Version count - how many versions are shown, for example Showing 1–8 of 20 versions.

• 5 Page numbers - move between pages of versions.

• 6 Trained versions - the group of versions created by training runs, with its count.

• 7 Baseline defaults - the group of models that ship with the platform. Fold or Show folds the group.

• Search matches name, model ID, stage, language, and status as you type. Press / anywhere on the page to jump to it, and Esc to clear it.

• Stage and Language narrow the list to one service or language.

• Sort orders versions newest first, oldest first, by highest version, or by name.

• Rows sets how many versions each page shows: 8, 16, 32, or 64.

The line under the toolbar says how many versions are shown and how many baseline defaults are folded. When a filter is active, a chip shows how many versions match and clears all filters.

Groups

• Trained versions — versions created by training runs. They come first.
• Baseline defaults — the 20 models that ship with the platform, one per service and language. They are protected from deletion. The group folds away once trained versions exist; click Show or Fold on its header.

Version rows

Click to enlarge

• 1 Model - the version name, its full model ID, and a button to copy the ID.
• 2 Status - for example Training, Completed, Active, Failed, or Stopped.
• 3 Delete - deletes the version after a confirmation.
• 4 Details toggle - expands or collapses the row.
• 5 Active · in use - why the version can’t be deleted.
• 6 Failed - the status with the first line of the failure reason.
• 7 Details - the model ID, dates, model files path, description, and failure reason.
• 8 Copy - copies the full model ID.

Each row shows:

Details

Click the arrow at the end of a row to expand it. Details show the full model ID with Copy, when the version was created, when training started and when it was last updated (server time), the model files path, the description, the full failure reason or training result, and why it can’t be deleted when that applies.

Activate a version

Click to enlarge

• 1 Active now - the model that is active for the language today.
• 2 After activation - the version that replaces it.
• 3 Activate - the row button that opens the dialog.
• 4 Cancel - closes the dialog without changes.
• 5 Activate - confirms the activation.

Activate appears on completed versions that aren’t active. It is disabled while the service trains or is busy. The dialog, titled Activate followed by the version name, shows the model that is Active now and the one that is active After activation for that language. Click Activate to confirm or Cancel. After activation the service loads the new model; the previous one stays in the list and can be activated again later.

Delete a version

The delete button (trash icon) removes a version and its model files after a confirmation; click Delete version to confirm. It can’t be undone.

• Default versions can’t be deleted; their rows say Default · protected.
• Active versions can’t be deleted; their rows say Active · in use. Activate another version for that language first.
• For a version still marked as training, the server first checks with the AI service that it isn’t actually training.

On small screens

Nothing is hidden on phones or tablets. The service tiles stack in pipeline order, the filters sit two by two, and each version becomes a card that keeps every column, its actions, and its expandable details.

Related pages

• AI settings — the training configuration each service reads when a run starts.
• Administration