Databricks Generative AI Engineer Associate Question 143
Select 3You are designing a metaprompt for a customer-facing chatbot powered by a large language model (LLM). The chatbot handles sensitive user data, such as credit card numbers, and must avoid hallucinating inaccurate responses. Which strategies should you include in your metaprompt to minimize hallucinations and prevent leaking private data?
- A
Instruct the model explicitly to avoid generating responses that include sensitive information unless explicitly provided by the user.
- B
Limit the model's output length to reduce the risk of hallucinations.
- C
Incorporate clear instructions to prioritize factual accuracy and cite sources where applicable.
- D
Provide examples of safe and unsafe responses within the metaprompt to guide the model's behavior.
- E
Encourage the model to generate creative responses to ensure engaging user interactions.
Show answer and explanation
Correct answers: A, C, D
Explanation
Minimizing hallucinations and preventing private data leaks require clear and explicit instructions within the metaprompt. Strategies like instructing the model to avoid sensitive information, emphasizing factual accuracy, and providing examples of acceptable behavior directly address these risks. However, focusing on creativity or limiting output length does not specifically tackle these issues and may inadvertently introduce new challenges.
- A. Correct.
Explicitly instructing the model to avoid generating sensitive information is crucial for preventing data leakage and ensuring user privacy. It sets clear boundaries for the model's behavior.
- B. Incorrect.
While limiting output length can be useful in some contexts, it does not directly address hallucinations or the risk of leaking private data.
- C. Correct.
Prioritizing factual accuracy and citing sources helps reduce hallucinations by guiding the model to focus on verifiable information.
- D. Correct.
Providing examples of safe and unsafe responses helps the model better understand the boundaries of acceptable behavior, reducing the likelihood of hallucinations and data leaks.
- E. Incorrect.
Encouraging creative responses might lead to hallucinations, as the model could prioritize creativity over accuracy or safety.