AI-102 Question 410
Select 3You are tasked with creating a solution that processes invoices using a custom document intelligence model in Azure. After uploading several sample invoice documents and labeling the fields, you train the model. Before publishing it, you want to ensure the model performs well on unseen data. Which steps should you take to validate the model before publishing?
- A
Split the dataset into training and testing datasets and evaluate the model using the testing set.
- B
Use the model's performance metrics, such as accuracy and recall, provided after training.
- C
Publish the model and test it on production data for validation purposes.
- D
Manually test the model by uploading a few unseen invoices and verifying the extracted data.
- E
Re-train the model using the same dataset to improve its performance.
Show answer and explanation
Correct answers: A, B, D
Explanation
To validate a custom document intelligence model before publishing, you need to ensure it performs well on unseen data. This involves splitting the data for training and testing, analyzing performance metrics like accuracy and recall, and performing manual tests with new documents. Testing directly in production or re-training with the same data does not provide reliable validation.
- A. Correct.
Splitting the dataset into training and testing datasets is a best practice to ensure the model generalizes well to unseen data. It allows you to measure its performance before publishing.
- B. Correct.
Using the provided performance metrics like accuracy and recall after training helps you analyze how well the model performed during testing.
- C. Incorrect.
Testing a model directly in production without prior validation is risky and not recommended. This approach can lead to poor user experiences if the model is not reliable.
- D. Correct.
Manually testing the model with unseen invoices is a good step to supplement automated testing. It helps verify the real-world applicability of the model.
- E. Incorrect.
Re-training the model with the same dataset without adding new data does not improve performance. Instead, it may lead to overfitting.