Databricks Generative AI Engineer Associate Question 137
Select 3You are deploying a large language model (LLM) in a customer service chatbot to handle user queries. During testing, you notice that the LLM occasionally produces harmful or inappropriate responses to sensitive topics. Which steps should you implement to establish guardrails and prevent negative outcomes?
- A
Utilize prompt engineering to reframe the input queries and steer the model's behavior.
- B
Integrate a content moderation layer to detect and filter harmful outputs before they are displayed.
- C
Disable logging and monitoring of the chatbot interactions to ensure user privacy.
- D
Implement fine-tuning of the LLM with domain-specific and ethical training data.
- E
Allow the LLM to operate without intervention, trusting its pre-trained behavior for all inputs.
Show answer and explanation
Correct answers: A, B, D
Explanation
Establishing guardrails for LLMs involves multiple strategies, including prompt engineering to guide behavior, content moderation to filter inappropriate outputs, and fine-tuning to align the model with ethical and domain-specific standards. Disabling monitoring or relying solely on pre-trained behavior are not effective strategies, as they either remove visibility into harmful patterns or fail to address inherent risks in the model's outputs.
- A. Correct.
Reframing input queries using prompt engineering can steer the LLM's behavior, reducing the likelihood of harmful or inappropriate responses.
- B. Correct.
A content moderation layer acts as a safeguard by detecting and filtering harmful outputs before they reach the user, ensuring safety.
- C. Incorrect.
Disabling logging and monitoring sacrifices the ability to track harmful patterns or troubleshoot issues, which is counterproductive for ensuring guardrails.
- D. Correct.
Fine-tuning the LLM with domain-specific and ethical training data helps align the model's behavior more closely with desired outcomes and reduces risks of harmful outputs.
- E. Incorrect.
Allowing the LLM to operate without intervention increases the risk of negative outcomes, as pre-trained models are not inherently aligned with all ethical or safety requirements.