Train

User permissions required: View Sources AND Review and label.

Overview

The main Train page provides useful information about the training done so far, the performance of the model, and a list of prioritized next best training actions to take same as the Validation page. It is a fully guided label training experience.

Train home page for a partially trained dataset

Note: The progress indicators give you further context while training the model with more granular validation feedback. These indicators encourage you to complete foundational training actions before focusing on specific performance factors. Building up a sufficient proportion of training data is important before focusing on model refinement.

Hover over the annotation progress areas to see additional performance information and underlying contributors. In the example below you can check the additional performance information once the foundational training actions are complete:

To train an action:

1. Select a training action to go to the specific training batch interface, for short, easy-to-consume training sessions.

   Depending on the recommended action, the number of messages or clusters of messages in the batch is 10, but it can vary.

   Batch training page for 'Shuffle' training
   

   

   
   
2. Apply the labels (and entities) to the message(s) on the screen.
3. Click Done. You can move onto the next message or cluster by clicking Next.
4. A summary of the training actions that you took, is displayed at the end of the batch. Select another recommended action to choose your next session.
   Summary of training actions completed during a training batch
5. Select another recommended action to choose your next session.

Note: The training that you complete triggers a re-training, and the next best actions also update as soon as possible. It's normal if more sessions of the same recommended action are required, and you might see the same actions on the page. You don't have to wait for new actions to appear, to jump into another similar session.

If you prefer to train without the platform's guidance, you can disable the Guided toggle icon and select which sessions to complete. For more details, check the Using Train without guidance enabled for labels section below.

How does it impact the model training process?

Train will further become the main place to complete all of your model training from start to finish, but some additional features are still in development (e.g. guided entity training). Right now, it's an add-on to the existing feature set, meaning that all of the functionalities you're used to can be used as-is, and you can train models as you usually do.

It is recommended that you use Train for a guided label training experience, and provide feedback to your UiPath Account Manager if encountering any issues or challenges.

How to use Train as part of model training going forward

Label training

Note: All of the pre-existing training modes are still available as they were via the Discover and Explore pages.

Training in Train:

• Guides you right from the moment you create a dataset with the next best actions to take to advance your label training - this includes uploading a taxonomy before you begin training
• Guides you through the usual steps covered elsewhere in this Knowledge Base for the model training process (check Overview), with the exception of recommending search
  • For an effective training mode, use the Search action sparingly, to provide the model with a limited set of initial examples for labels that don't have enough training data yet. To use this action, go to Discover, Explore, or by temporarily disabling the guidance in Train (check the Using Train without guidance enabled for labels section for more details).
• Provides need to know performance feedback in the main page and through its recommendations. If you need detailed feedback on model performance, go to the Validation page.

Train page for a completely new dataset

Note: Hover over the annotation progress areas to see the additional progress indicators.

Entity training

Toggle between training labels and entities on the train tab if you have entities enabled on your dataset.

Toggle to switch between training labels and entities on the Train tab

Note: Similar to training labels, all of the pre-existing training modes for entities are still available as they were via the Explore page.

Training entities in Train:

• Guides you right from the moment you create a dataset with the next best actions to take to advance your entity training.
• Guides you through the usual steps covered elsewhere in this Knowledge Base for training entities during the model training process.
• Provides need to know performance feedback in the main page and through its recommendations. If you need detailed feedback on entity performance, go to the Validation, then Entity Validation pages.
• During the beginning of the model training process, if the platform doesn't have enough examples of entities to learn from - it will recommend shuffle by default. Once you provide enough examples, it will recommend more targeted training for specific entities.

Train page for entities on a trained dataset

Using Train without guidance enabled for labels

The default setting for the Train page is to have platform guidance enabled, as this is our recommendation.

If you're a confident model trainer and you know the actions that you already want to take, you can disable the guidance, using the toggle in the top right-hand of the page:

Toggle for enabling/disabling platform guidance

Unguided Train interface

The platform will still highlight the phase of training it thinks is most appropriate. You can find the usual training actions within each phase, and you'll be able to target specific labels as needed (see below).

Teach label selector using unguided Train interface