NCA-GENL Question 67
Select 3You are working with a dataset to train a generative AI large language model (LLM). During the preprocessing step, you notice that the dataset contains missing values, inconsistent text formats, and irrelevant columns. Which of the following actions would be most appropriate to ensure high-quality data for training the LLM?
- A
Impute missing values using techniques like mean, median, or a placeholder token depending on the context.
- B
Normalize inconsistent text formats, such as converting all text to lowercase and standardizing date formats.
- C
Remove irrelevant columns and rows that do not contribute to the target output or model goals.
- D
Ignore the missing values and inconsistent formats since LLMs are robust to noisy data.
- E
Add synthetic data to replace missing values without considering the original data distribution.
Show answer and explanation
Correct answers: A, B, C
Explanation
To train an effective generative AI LLM, it is critical to preprocess the dataset by addressing missing values, standardizing text formats, and removing irrelevant information. These steps ensure the data is clean, consistent, and meaningful, which is essential for high-quality model training. Ignoring data issues or introducing synthetic data without careful consideration can result in suboptimal model performance and biased outputs.
- A. Correct.
Imputing missing values is a standard data preprocessing step to address gaps in the dataset. For text-based data, replacing missing values with a placeholder token ensures consistency and prevents disruption in training.
- B. Correct.
Normalizing inconsistent text formats helps eliminate variability caused by case sensitivity or non-standard representations, which improves the overall quality of the input data for training.
- C. Correct.
Removing irrelevant columns and rows reduces noise in the data and helps focus the model on learning from meaningful and relevant information.
- D. Incorrect.
Ignoring missing values and inconsistent formats can lead to poor model performance, as LLMs require clean and structured data for effective training.
- E. Incorrect.
Adding synthetic data without considering the original distribution can introduce bias or unrealistic patterns, negatively impacting model performance.