Databricks Generative AI Engineer Associate Question 273
Single answerYou are evaluating a generative AI model deployed on Databricks to summarize customer feedback. Despite achieving high BLEU scores during training, users report that the summaries often miss important details or include fabricated content. Which approach would best help you evaluate and monitor the model’s real-world performance?
- A
Use human evaluation to assess the relevance and factual correctness of the summaries.
- B
Increase the size of the training dataset to improve the model's accuracy.
- C
Monitor only quantitative metrics like BLEU and ROUGE scores on test data.
- D
Deploy the model to production without further evaluation since the training metrics are strong.
Show answer and explanation
Correct answer: A
Explanation
Quantitative metrics like BLEU and ROUGE are useful but do not fully capture the quality of generative AI outputs, especially for subjective tasks like summarization. Human evaluation provides deeper insights into real-world issues such as relevance, factual accuracy, and hallucination, making it essential for effective evaluation and monitoring.
- A. Correct.
Human evaluation is critical for assessing real-world applicability, as it can uncover issues like factual inaccuracies or missing information that automated metrics may not detect.
- B. Incorrect.
Increasing the training dataset size may improve accuracy, but it does not directly address the issue of evaluating real-world performance or monitoring after deployment.
- C. Incorrect.
Relying solely on metrics like BLEU and ROUGE may not capture qualitative issues such as relevance or hallucination in outputs.
- D. Incorrect.
Deploying without further evaluation ignores the feedback from users and could lead to poor performance in real-world scenarios.