Databricks Generative AI Engineer Associate Question 134
Select 3You are deploying a generative AI model for a customer support chatbot. To ensure the model does not generate harmful or biased responses, you plan to implement guardrails. Which of the following actions are effective strategies to implement guardrails for this use case?
- A
Use prompt engineering to include explicit instructions like 'Avoid providing harmful or biased responses.'
- B
Integrate a content moderation API to filter the chatbot's responses before delivering them to users.
- C
Disable logging of user conversations to reduce system overhead.
- D
Implement a reinforcement learning from human feedback (RLHF) fine-tuning process to align the model with desired behaviors.
- E
Rely solely on the pre-trained model's default behavior without additional adjustments.
Show answer and explanation
Correct answers: A, B, D
Explanation
Implementing guardrails for generative AI models involves multiple strategies to reduce risks and ensure safe outputs. Effective approaches include prompt engineering to guide behavior, using moderation APIs for additional filtering, and fine-tuning with RLHF to better align the model with desired outcomes. These strategies work together to mitigate harmful or biased outputs while improving the system's reliability and safety.
- A. Correct.
Including explicit instructions in the prompt is an effective way to guide the model's behavior and reduce the likelihood of generating harmful or biased responses.
- B. Correct.
A content moderation API can act as a secondary filter to catch any inappropriate or harmful outputs from the generative AI model, ensuring safer interactions.
- C. Incorrect.
Disabling logging of user conversations does not directly contribute to implementing guardrails, and it might hinder the ability to audit or improve the system.
- D. Correct.
Fine-tuning the model using RLHF allows you to align the model's outputs with human values and desired ethical standards, making it a powerful approach for implementing guardrails.
- E. Incorrect.
Relying solely on the pre-trained model's default behavior is risky because it may not align with specific organizational needs, ethical considerations, or safety requirements.