NCA-GENL Question 127
Select 3You are tasked with conducting data analysis for fine-tuning a large language model (LLM) under the supervision of a senior team member. During the initial data exploration, you notice that the dataset contains a mix of labeled and unlabeled data, as well as some duplicate entries. What should you prioritize to ensure the data is ready for fine-tuning?
- A
Remove duplicate entries from the dataset.
- B
Ensure the labeled data is properly balanced across categories.
- C
Ignore unlabeled data as it cannot be used for fine-tuning.
- D
Work with the senior team member to define preprocessing steps for both labeled and unlabeled data.
- E
Focus only on labeled data and discard any unused features.
Show answer and explanation
Correct answers: A, B, D
Explanation
Conducting data analysis for fine-tuning LLMs requires a structured approach under supervision. Key priorities include eliminating duplicate data, ensuring balanced labeled data to prevent model bias, and defining preprocessing workflows in collaboration with senior team members. Ignoring unlabeled data or discarding unused features could result in missed opportunities for improving model performance or utilizing the full dataset effectively.
- A. Correct.
Removing duplicate entries is crucial to avoid bias or overrepresentation of certain data points, which can negatively impact model performance.
- B. Correct.
Ensuring the labeled data is balanced across categories helps prevent the model from becoming biased toward over-represented labels.
- C. Incorrect.
Unlabeled data can still be useful for tasks like semi-supervised learning or pretraining, so ignoring it entirely may not be the best approach.
- D. Correct.
Collaborating with the senior team member to define preprocessing steps ensures that the data preparation aligns with the overall project goals and methodologies.
- E. Incorrect.
Focusing only on labeled data and discarding unused features may lead to a loss of potentially valuable information, especially if those features could be relevant during preprocessing or training.