TDM for Classification Models

Training Data Management for Classification is where you prepare the data for a Semi-structured Classification model. This model is also called an NLC model. It selects the Semi-structured or Additional layout for each page in a submission. For more information, see Semi-structured Document Classification.

Before you use TDM: A release contains the layouts that the model must recognize. You collect documents that represent each layout.

In TDM: You assign documents to layouts. You also add documents that the model must not match.

After you prepare the data: You train a model version, review its results, and deploy the version that you want to use.

This article explains the model-version controls in Model History and the document controls in the Training Data tab.

Accessing this feature
Your access to the feature described in this article depends on your license package and pricing plan.
To learn which features are available to your organization and how to add more, contact your Hyperscience representative.

Before you start

Make sure that the release contains the correct Semi-structured or Additional layouts.
Collect different examples of the documents that you expect to process.
Collect examples that the model must ignore. These are called Excluded documents.

Use processed submissions as training data

Enable Send documents to Training Data Management in Administration > System Settings. This setting is off by default.

Open TDM for Classification

Go to Models.
Open the Classification tab.
Select a model name.

The Classification tab in the Models section, showing the Classification models table

You can also open a model from its release. Go to Library > Releases. Open the release, then select View Model in the Automatic Document Classification card.

The Automatic Document Classification card on a release's page, with the View Model button

The Classification models table shows the model name, compatible releases, status, and last deployment date. Use Filter by release to show models for one release.

Importing duplicate Classification models

In v43.2 and later, importing a Classification model with a UUID that already exists returns an error instead of creating a duplicate version. The error identifies the UUID and notes that the existing model may be in the Archive view.

Use TDM for Classification

Follow this order when you prepare and train a Classification model.

Step	What to do	Result
1. Add data	Add documents for each layout. Add Excluded documents for pages that the model must ignore.	The model has examples of correct and incorrect matches.
2. Review data	Check each layout assignment. Remove incorrect or duplicate documents. Set usage rules and tags when needed.	The training data is correct and easy to manage.
3. Check the page count	Check the Training Data Summary and Training Data Health cards.	You know if the model is ready to train.
4. Train and review	Run training. Review projected automation and any Model Validation Tasks (MVTs).	You have a new model version and its results.
5. Deploy or improve	Deploy the model version, or update the data and train again.	The selected version becomes the live model.

Training data requirements

You manage whole documents in TDM. The model trains and predicts at the page level. Therefore, TDM uses page counts to check if the model is ready.

Minimum: 10 pages for each layout.
Recommended: 120 pages for each layout.
Maximum per layout: 5,000 pages.
Maximum per model: 100,000 pages.

Excluded pages are required when the release has one Semi-structured or Additional layout. They are optional when the release has two or more layouts. Add them when the model must learn to ignore similar pages.

Do not add the same document more than once.

Overview tab

Use the Overview tab to review model performance, model versions, and release compatibility.

Projected Automation chart

The chart shows projected automation for the live model. A higher target accuracy usually gives lower automation. A lower target accuracy usually gives higher automation. The chart also shows the margin of error.

For more information, see Accuracy, Automation, and Evaluating Model Training Results.

The Projected Automation chart in the Overview tab

Model History table

The table shows each model version. It includes the name, UUID, creation date, trainer version, source, projected automation, documents used for training, and last deployment date.

You can search, filter, sort, and choose which columns to show. You can also change the test target accuracy.

The Actions menu can include these options:

Deploy: Make this version live.
Undeploy: Stop using the live version.
Reject Candidate: Reject a candidate version.
Download: Download the version for use in another instance.
Archive: Starting in v43.2, move the version from Model History to the Archive view.
[v44 NEW] Rename: Starting in v44.0, change the displayed name of the version. Renaming does not create a new version.

[v44 NEW — section start]

Rename a model version

In v44.0 and later, you can rename a model version in the Model History table. Renaming changes only the displayed name. It does not create a new version.

In the Model History table, find the version to rename.
Open its three-dot Actions menu and select Rename. [Verify with Design: whether Rename is in the Actions menu or inline on the name, and the field label in the dialog.]
Enter the new name, and confirm.

When you download a renamed version and import it into another instance, its legacy name is preserved. [Verify with Engineering: whether the imported version shows the new name, the legacy name, or both, and where the legacy name is visible.]

[v44 NEW — section end]

Archive and unarchive model versions

In v43.2 and later, archived versions are removed from the active Model History view but remain available in the Archive view. You can archive a version only when it is neither live nor a candidate. Undeploy a live version or reject a candidate version first.

In the Model History table, find the version to archive.
Open its three-dot Actions menu and select Archive.
Confirm the action.

To restore a version, open the table options menu and select View Archive. Open the archived version’s three-dot menu and select Unarchive. The version returns to Model History as inactive; it is not deployed automatically. Select View Active to return to the active view.

Note: You can permanently delete an archived version. Deletion cannot be undone.

View and copy a model UUID

In v43.2 and later, the UUID column is hidden by default. In Model History, open the table options menu next to Filter, select Manage columns, enable UUID, and save. The column shows a shortened identifier. Hover over it to view the complete UUID, or select it to copy the complete value to your clipboard.

The Model History table with filtering, searching, and column management options

Model Compatibility table

This table lists the releases that can use the model. It shows each release name, creation date, and status. Select a column heading to sort the table. For release statuses, see What is a Release?.

Training Data tab

Use the Training Data tab to check and manage the documents used for training.

Training Data Summary

This card shows if the training data meets the page requirements.

Status	Meaning
Requirements Not Met	At least one layout has fewer than 10 pages. You cannot start training.
Not Optimized	Each layout has at least 10 pages, but one or more layouts have fewer than 120 pages. You can start training.
Ready To Train	Each layout has at least 120 pages.

The card shows eligible layouts, page counts, document counts, and Excluded data counts. Excluded pages do not appear in the main page count when they are optional.

The Training Data Summary card

Training Data Health

This card shows the page count for each layout. It also shows the required and recommended counts. Select Add Training Data to add documents.

The Training Data Health card, showing per-layout page counts

Training Data table

This table lists the documents that can be used for training.

The Training Data table

Columns: Document ID, File Name, Pages, Layout, Usage Rule, Source, Scheduled Deletion, and Tags.

Point to the preview icon next to a Document ID to preview the document. Select the icon to keep the preview open.
Older documents can have no file name. Imported older data shows N/A.
Source can be Upload, Processing, or QA. QA applies to legacy documents from v37 or earlier.

Filters: Layout, Usage Rule, Source, Scheduled Deletion, Has Anomalies, and Tags.

Actions: Edit the layout or usage rule, export all or selected rows, or delete selected rows.

[v44 NEW — section start]

Edit or delete documents in a filtered table

In v44.0 and later, when you filter the Training Data table, the edit and delete actions apply only to the documents that you select in the filtered results. Documents that the filters hide are not changed.

To apply an action to every document that matches the filters:

Apply the filters.
Select the checkbox in the table header.
Select the option that selects all documents that match the filters. [Verify with Design: the exact label of this control and where it appears.]
Open the Actions menu, and select the edit or delete action.

[Verify with Product: whether this also applies to the Excluded Training Data table.]

[v44 NEW — section end]

Export training data as CSV

To export a subset, select its rows in the Training Data table. To export the entire table, do not select individual rows.
Open the table’s Actions menu. Under Export, select All training data as CSV or Selected rows as CSV.
When the export is ready, use the download link in the Notifications menu.

A CSV export option is unavailable if it would include documents that are still loading. All training data as CSV is unavailable while any document in that table is loading. Selected rows as CSV remains available for loaded rows, even if another document is loading, but is unavailable when the selection includes a loading document. Hover over an unavailable option to see why. Wait for loading to finish or select only ready rows.

The same CSV export behavior applies to the Excluded Training Data table.

Usage rules

Value	How the system uses the document
Always	Uses the document in future training. Data-deletion settings do not delete it.
Auto	Can use the document until data-deletion settings delete it.
Never	Does not use the document. Data-deletion settings can delete it. Documents from Supervision or QA always use this value.
Loading	The document is still loading. The value changes to Auto after loading finishes.

Anomalies

A high-confidence error occurs when the model is confident but the layout is wrong. TDM marks these documents as anomalies. Complete the related Model Validation Tasks (MVTs) to check and correct them. For more information, see Model Validation Tasks.

Excluded Training Data table

This table contains documents that the model must not match to a layout. It uses the same columns, filters, and actions as the Training Data table.

The Excluded Training Data table

For CSV export steps, see Export training data as CSV.