NCA-GENL Question 129
Select 3You are assisting your senior team member in analyzing a dataset for training a generative AI large language model (LLM). The dataset contains text samples from multiple domains, but some entries have missing values or encoding issues. What steps should you take under supervision to ensure the dataset is prepared for training?
- A
Identify and remove entries with missing values or encoding issues, as they may introduce errors during training.
- B
Perform data augmentation to artificially increase the size of the dataset before cleaning it.
- C
Generate a report summarizing the distribution of text lengths, domain representation, and missing value percentages.
- D
Normalize the dataset by converting all text to lowercase and removing special characters, if advised by the senior team member.
- E
Directly feed the raw dataset into the training pipeline to save time, assuming the model will handle data inconsistencies.
Show answer and explanation
Correct answers: A, C, D
Explanation
Conducting data analysis for training LLMs involves identifying and addressing data quality issues, understanding dataset characteristics, and normalizing the data as needed. Steps like removing problematic entries, generating comprehensive reports, and normalizing text ensure the dataset is suitable for training. Data augmentation and feeding raw data are inappropriate at this stage.
- A. Correct.
Removing entries with missing values or encoding issues is a critical step to ensure data quality and prevent issues during training.
- B. Incorrect.
Data augmentation is typically performed after cleaning the dataset, not before. Cleaning ensures that the augmentation process is applied to high-quality data.
- C. Correct.
Generating a report is an important step to understand the dataset's characteristics and identify any imbalances or issues that need to be addressed.
- D. Correct.
Normalization can be necessary to standardize the data and make it suitable for the model, but it should be done under the guidance of the senior team member.
- E. Incorrect.
Feeding raw, uncleaned data into the training pipeline can lead to poor model performance and errors, as LLMs require high-quality, preprocessed data.