Project versioning

Note

This article applies to the following custom features in Azure AI Language:

• Conversational language understanding
• Custom text classification
• Custom NER
• Orchestration workflow

Building your project typically happens in increments. You may add, remove, or edit intents, entities, labels and data at each stage. Every time you train, a snapshot of your current project state is taken to produce a model. That model saves the snapshot to be loaded back at any time. Every model acts as its own version of the project.

For example, if your project has 10 intents and/or entities, with 50 training documents or utterances, it can be trained to create a model named v1. Afterwards, you might make changes to the project to alter the numbers of training data. The project can be trained again to create a new model named v2. If you don't like the changes you've made in v2 and would like to continue from where you left off in model v1, then you would just need to load the model data from v1 back into the project. Loading a model's data is possible through both the Language Studio and API. Once complete, the project will have the original amount and types of training data.

If the project data is not saved in a trained model, it can be lost. For example, if you loaded model v1, your project now has the data that was used to train it. If you then made changes, didn't train, and loaded model v2, you would lose those changes as they weren't saved to any specific snapshot.

If you overwrite a model with a new snapshot of data, you won't be able to revert back to any previous state of that model.

You always have the option to locally export the data for every model.

Data location

The data for your model versions will be saved in different locations, depending on the custom feature you're using.

Custom NER

In custom named entity recognition, the data being saved to the snapshot is the labels file.

Custom text classification

In custom text classification, the data being saved to the snapshot is the labels file.

Orchestration workflow

In orchestration workflow, you do not version or store the assets of the connected intents as part of the orchestration snapshot - those are managed separately. The only snapshot being taken is of the connection itself and the intents and utterances that do not have connections, including all the test data.

Conversational language understanding

In conversational language understanding, the data being saved to the snapshot are the intents and utterances included in the project.

Next steps

Learn how to load or export model data for:

• Conversational language understanding
• Custom text classification
• Custom NER
• Orchestration workflow