NCA-GENL Question 194
Select 3A company is deploying a large language model (LLM) and wants to ensure the system is ethical, energy-conscious, and reliable in its outputs. Which of the following actions would best align with these goals?
- A
Implement dataset filtering to remove biased or inappropriate training data.
- B
Utilize model compression techniques to reduce energy consumption during inference.
- C
Allow the model to generate uncensored outputs to ensure transparency.
- D
Conduct regular audits of the model's predictions and behavior for fairness and reliability.
- E
Train the model using only publicly available data without verifying its source or quality.
Show answer and explanation
Correct answers: A, B, D
Explanation
To create ethical, energy-conscious, and reliable LLMs, it is necessary to filter training datasets to remove biases, apply techniques like model compression to reduce energy consumption, and perform regular audits to ensure fairness and reliability. Actions that disregard data quality or promote uncensored outputs without safeguards can undermine these goals.
- A. Correct.
Filtering biased or inappropriate training data is essential to ensure the model produces ethical and fair outputs. Using poor-quality datasets can lead to unreliable and unethical results.
- B. Correct.
Model compression techniques, such as pruning or quantization, help reduce energy consumption during inference, making the system more energy-conscious without significantly affecting performance.
- C. Incorrect.
Allowing the model to generate uncensored outputs might increase transparency but could also lead to unethical or harmful outputs, which is against the goal of creating ethical systems.
- D. Correct.
Regular audits ensure the model aligns with fairness and reliability standards, helping identify and mitigate issues in predictions or behavior.
- E. Incorrect.
Using publicly available data without verifying its source or quality can introduce unreliable or biased information into the model, violating ethical and reliability standards.