NCA-GENM Question 56
Select 3You are tasked with preparing a large, multimodal dataset for training a generative AI model. The dataset includes text descriptions, images, and audio files. During the data preparation process, you notice that some images are mislabeled, some audio files are corrupted, and several text descriptions contain irrelevant metadata. Which steps should you prioritize to ensure the dataset is clean and ready for training?
- A
Manually inspect and relabel the mislabeled images to ensure accuracy.
- B
Remove corrupted audio files to prevent errors during the training process.
- C
Augment the dataset by duplicating text descriptions to increase the volume of data.
- D
Remove irrelevant metadata from text descriptions to maintain data quality.
- E
Directly train the model on the dataset without any modifications to save time.
Show answer and explanation
Correct answers: A, B, D
Explanation
In multimodal generative AI, the quality of the dataset directly impacts the performance of the model. Cleaning and transforming the data by relabeling mislabeled images, removing corrupted files, and eliminating irrelevant metadata ensures that the input data is accurate, consistent, and useful for training. Ignoring these steps or adding unnecessary augmentations can lead to poor model performance or biases.
- A. Correct.
Manually inspecting and relabeling mislabeled images is crucial to ensure high-quality data, as incorrect labels can negatively impact the model's learning process.
- B. Correct.
Removing corrupted audio files prevents the model from encountering errors or learning from incomplete or invalid data.
- C. Incorrect.
Duplicating text descriptions does not improve the quality of the dataset and may introduce biases or overfitting during training.
- D. Correct.
Removing irrelevant metadata from text descriptions is necessary to maintain data clarity and consistency, reducing noise in the dataset.
- E. Incorrect.
Training the model without any modifications would result in poor model performance, as the dataset contains significant issues that need to be addressed.