Magnolia AI Training

This module of Magnolia CMS provides a tool to fine-tune Large language models like GPT using as input dataset contents in Magnolia CMS.

Features

• Integration with the API of OpenAI and Azure OpenAI
• Export of contents in Magnolia to datasets for fine-tuning models of Open AI and Azure OpenAI

Setup

1. Add maven dependency to your bundle of Magnolia

   <dependencies>
   <dependency>
       <groupId>org.formentor</groupId>
       <artifactId>magnolia-ai-training</artifactId>
       <version>0.0.1</version>
   </dependency>
   </dependencies>

2. Activate and configure Open AI or Azure OpenAI

OpenAI

• Specify host of OpenAI and model to fine-tuned in the configuration of the module.

config:/formentor-ai-training/openAI

• Specify API key in the secret openai/api-key of Password Manager.

Azure OpenAI

• Specify apiVersion and url of the Azure resource and the model to be fine-tuned in configuration of the module.

config:/formentor-ai-training/azure

• Specify the API key in the secret azure/api-key of Password Manager.

Usage

The module ai-training provides the command ModelTrainerCommand to create fine-tuned models

Input parameters of ModelTrainerCommand

• modelName to specify the prefix of the fine-tuned model created.
• workspace to specify the workspace of the text contents to be used to build the training data for fine-tuning.
• root to specify the root path of the nodes to be used to build the examples of the datataset.
• nodeType to specify of the node type of the nodes to be used to build the training data.
• propertiesAsPrompt to specify the properties to be used to build the prompt in the training data. In case of linked fields specify the target worspace and target property
• propertyAsCompletion to specify the property used as completion in the examples of the training data.

Example

cm = info.magnolia.commands.CommandsManager.getInstance()
command = cm.getCommand('formentor','FineTuneModelCommand')

/**
 *
 * This example fine-tunes the model using as dataset:
 * INPUT: properties "tourTypes", "category" and "displayName" of nodtypes "mgnl:content" in the workspace "tours" from path "/magnolia-travels" 
 * COMPLETION: property "body"
 *
 * The fine-tuned model will be trained for the task of writing description of tours from the factsheet
 */
propertyPromptTourTypes = new com.formentor.magnolia.ai.training.domain.PropertyPromptValue()
propertyPromptTourTypes.setPropertyName('tourTypes')
reference = new com.formentor.magnolia.ai.training.domain.PropertyPromptValue.Reference()
reference.setTargetWorkspace("category")
reference.setTargetPropertyName("displayName")
propertyPromptTourTypes.setReference(reference)

propertyPromptLocation = new com.formentor.magnolia.ai.training.domain.PropertyPromptValue()
propertyPromptLocation.setPropertyName('location')

command.setWorkspace('tours')
command.setModelName('magnolia-tours')
command.setRoot('/magnolia-travels')
command.setNodeType('mgnl:content')
command.setPropertiesAsPrompt([propertyPromptTourTypes, propertyPromptLocation])
command.setPropertyAsCompletion(['body'])

command.execute(new info.magnolia.context.SimpleContext())