5.4. Training Models
Training Models
Extractor Processors have four types of models that can be trained that fall under three categories found on the Build Tab below the Dataset Overview under Create New Versions. The three categories are Call Foundation Model, Fine Tuning, and Train a Custom Model. All three categories can be accessed from the Build tab in the Create New Versions section (Red Border) below the Dataset Overview section by clicking on the category's "Create New Version" buttons (Blue Arrows). Additionally, the Train a Custom Model category can be accessed from the Manage Dataset Page by clicking on the "Train New Version" button (Red Arrow). Each button will open a right side panel used to start training a model. If a less important Label is preventing Model Training because it doesn't meet the requirements it is recommended to disable the Label in the Schema to remove the Label from the Model training requirements, this will not remove Labels in the document files. When the requirements are met a new Model can be trained, it is suggested to use sequential naming to differentiate between multiple Model Training Phases.
In order to train a model all enabled Labels must meet the minimum required number of Documents with each Label in the Training and Testing Sets. Clicking on the "View Label Stats" button (Red Arrow) shows the distribution of each Label.
The Field Labels and Table Labels that do not meet the requirements are listed at the top of the Label List and flagged with a red warning. Note: If a Table Label meets the requirements but one of its "Child" Field Labels do not meet the requirements, the Table Label will be listed in its alphabetical position in the Label List and the "Child" Field Label will be listed at the "Child" Field Labels within the Table Label and flagged with a red warning.
When all active Labels meet the requirement, but not all meet the recommended number of Documents the message "Your dataset does not meet the recommended training criteria. You can still train a new version, but your model's quality may be affected." with an orange warning flag will be shown above the Label List. If the recommended number of Training Documents is not met an orange warning flag will be next to "Each label exists on 50 training documents (minimum: 10)". If the recommended number of Testing Documents is not met an orange warning flag will be next to "Each label exists on 50 test documents (minimum: 10)".
Call Foundation Model
Call Foundation Model creates a Model from Google's Pretrained Foundation Models with zero training using only the Schema for tuning and has the requirement that at least one Field is defined in the Schema. When creating a Foundation Model the Model's Base Version is selected from a dropdown list of the available Pretrained Foundation Models.
The OGRRE Workflow does not make use of Call Foundation Model.
Fine Tuning
Fine Tuning creates a Model from Google's Pretrained Foundation Models by using the Labeled Documents and the Schema for tuning and recommends that each Field is present in 50 Documents in the Training Set and 50 Documents in the Testing Set of the Labeled Documents while requiring only that each Field is present on a minimum of 1 Document in each Document Set. When creating a Fine Tuning Model the Model's Base Version is selected from a dropdown list of the available Pretrained Foundation Models. Fine Tuning has two Tuning Setting called Train Step and Learning Rate Multiplier the OGRRE Workflow uses the default Tuning Setting parameters.
Train a Custom Model
Train a Custom Model creates a Model without the use Google's Pretrained Foundation Models relying only on the Labeled Documents and the Schema to create the Model. Custom Models can be either Model Based or Template Based selected with the radio buttons under "Model Training Method".
Template Based
Template Based Custom Models require that each Field is present in 3 Documents in the Training Set and 3 Documents in the Testing Set of the Labeled Documents.
The OGRRE Workflow does not make use of Template Based Custom Models because of their inability to handle any variation.
Model Based
Model Based Custom Models recommend that each Field is present in 50 Documents in the Training Set and 50 Documents in the Testing Set of the Labeled Documents while requiring that each Field is present in 10 Documents in the Training Set and 10 Documents in the Testing Set of the Labeled Documents.
OGRRE Workflow
The OGRRE Workflow sets out that during each Model Training Phase a new Fine Tuning Model is trained from each of the available versions of Google's Pretrained Foundation Models along with one new Model Based Custom Model. As a result, the OGRRE Workflow follows the requirements set in place for training a Model Based Custom Model of a minimum of 10 Documents in the Training Set and 10 Documents in the Testing Set of the Labeled Documents for each Field before training any Model. Prior to initiating the Model Training Phase any Fields that do not meet the minimum requirement are disabled.