Document Understanding Modern Projects User Guide

DELIVERY:

Last updated Nov 20, 2025

Retrain extractors

You can use documents that were validated in Validation Station to further improve the performance of your models.

You can retrain documents that were processed using the following activities:

UiPath.DocumentUnderstanding.Activities: all documents that were processed using this activity package and were validated in Validated Station are collected automatically and can be used for retraining.
Note: If the validation task resides in a different environment than the Document Understanding project, the processed documents are not collected and available for retraining.
UiPath.IntelligentOCR.Activities (starting with version 6.25.0-preview): to retrain documents processed using this activity package, use the Document Understanding Project Extractor Trainer activity in your workflow. This allows documents to be collected for retraining purposes.

When there is at least one document collected, the Exceptions for review button is available for the corresponding document type within the Build section.

Note: Collected documents are not automatically included in the training set. You need to review the documents and confirm their addition in the training set to retrain your model.

Exceptions for review

Follow the steps from this procedure to fine-tune your model using documents from Validation Station.

Note: Documents collected for exceptions are stored for a period of 30 days, after which they are automatically deleted.

Documents are not collected for validation tasks with a duration greater than 7 days.

Select the Exceptions for review button.
Check the exception documents from the Exceptions for fine-tuning menu.
The following information is available for each document:
- File name: the file name of the document containing an exception.
- Status: the status of the document.
- Pages: the number of pages that the document contains.
- Project version: the project version that contains the document.
- Processed date: the date when the document was processed.
- No of extracted fields: the number of extracted fields for the document containing an exception.
- No of corrected fields: the number of modified fields during validation step.
- Validator name: the username of the person who validated the document.
Select the document you want to use for fine-tuning from the list.
The following information is available for each document:
- All extracted fields are displayed, along with their model confidence. If the confidence value is N/A, this indicates that the field was not automatically extracted and was instead manually added by the validator.
- Corrected fields are marked with a yellow point next to extracted field name.
- The following information is available for all corrected fields:
  - Predicted value: the value predicted by the model.
  - Corrected value: if changed manually, the value after validation. If the document type was not changed after validation, the value will be N/A.
  - Reference: the original value that is highlighted on the document. This value is used for annotation if you decide to use the document for fine-tuning.
- You can easily sort by confidences to check fields with low confidences without corrections.
- You can filter by corrected fields to only focus on them.
Choose Annotate if you want to further annotate this document or choose Use for fine-tuning if you want to use this document to retrain your model.
- Use for fine-tuning:
  - Choose this when the document is a good example for the model fine-tuning and all fields are correctly referenced within the document.
  - The document is imported in the training with the Exception tag and all annotations confirmed. The document will be used for fine-tuning.
- Annotate:
  - Choose this the document is a good example for the model fine-tuning, but there are some validation errors, for example not all fields are correctly referenced wihtin the document and further corrections are needed.
  - The document is imported in the training set with the Exception tag and all annotations unconfirmed. Annotations must be confirmed in the Build section in order for the document to be used to fine-tune the model.
  - If you choose to further annotate your document, check the Annotate documents page for more information on how to annotate documents.
- Exclude:
  - Choose this when the document is not a good example and you want to remove it from the list of exceptions for review and you do not need to review it in future review sessions.
  - You can change the status of the document to revert the changes.
Figure 1. Selected document used for fine-tuning or annotation

Once you selected all your documents for fine-tuning, your model is retrained using the new data from Validation Station.

You can create a new project version and use the Compare model feature from the Measure section to compare the performance of your model.

Exporting and importing re-training candidates

Document Understanding allows you to export re-training candidates from one environment and import them into another.

Exporting re-training candidates

You can export re-training candidates directly from the Exceptions for review page by clicking the Export button.

Note: Click View exports to access the list of exported files.

Importing re-training candidates

Importing is done from the Exceptions for review page. Imported documents are displayed in a similar manner to documents automatically collected from Validation Station, with the status To review.

Imported documents are not added directly to the training dataset. You must review and confirm them before they can be included in a retraining dataset.

On this page

Exceptions for review
Exporting and importing re-training candidates
Exporting re-training candidates
Importing re-training candidates

Was this page helpful?

PREVIOUSClassify documents automatically

NEXTTrain a classifier

Support and Services

Get The Help You Need

UiPath Academy

Learning RPA - Automation Courses

UiPath Forum

UiPath Community Forum

Trust and Security

Cookies Policy