AI-102 Question 409
Single answerYou are designing a solution using Azure Form Recognizer to extract structured data from invoices. You want to build a custom model to process these documents. Which sequence of steps should you follow to train, test, and publish the custom document intelligence model?
- A
Upload labeled training data to a storage account, train the model using labeled data, test the model with a validation dataset, and publish the model.
- B
Upload unlabeled training data to a storage account, train the model using prebuilt templates, test the model using the Quick Test feature, and publish the model.
- C
Upload labeled training data to a storage account, train the model using labeled data, evaluate the model performance using the Analyze API, and publish the model.
- D
Upload unlabeled training data to a storage account, train the model using a general prebuilt model, test the model with a validation dataset, and publish the model.
Show answer and explanation
Correct answer: C
Explanation
To train, test, and publish a custom document intelligence model in Azure Form Recognizer, you need to upload labeled training data to a storage account, train the model using this labeled data, evaluate the model's performance using the Analyze API to ensure it meets accuracy requirements, and then publish the model for use in production. This ensures the model is trained for specific document types and performs effectively.
- A. Incorrect.
This is incorrect because while it mentions labeled training data and testing, it misses the key step of evaluating the model's performance using the Analyze API before publishing.
- B. Incorrect.
This is incorrect because using prebuilt templates is not applicable when training a custom model, and Quick Test is not a replacement for proper testing and evaluation.
- C. Correct.
This is correct because it outlines the correct process: uploading labeled training data, training the model, evaluating its performance using the Analyze API, and then publishing the model.
- D. Incorrect.
This is incorrect because using unlabeled training data and general prebuilt models does not apply when creating a custom model for specific document processing tasks.