Section 3: Techniques to improve generative AI model output約20% · 5 topics

Q1. For a chat answering internal policy questions, you want users to be able to verify the basis of answers. Which method fits best?

1. Ground answers in the internal policy documents and show sources
2. Fine-tune the model on policy documents so it memorizes them
3. Lower temperature to reduce variation in answers
4. Enable Grounding with Google Search to supplement answers with web information

Q2. What is the most appropriate countermeasure to the “bias” limitation of foundation models?

1. Evaluate from diverse viewpoints, and revisit data or operations if bias is found
2. Models are neutral, so there's no need to check for bias
3. Bias is unrelated to training data, so don't check the data
4. Hide biased output so users don't notice

Q1. Which are the best example metrics to track in “continuous monitoring” of generative AI models?

1. Length of the model's name and color of its logo
2. Employees' commuting time and office days
3. Age of the data center building
4. Answer accuracy, satisfaction and response time

Q2. Why is “version management” of generative AI models important?

1. Version management is needed only for image generation
2. Model updates change output, so record versions to compare and roll back
3. Higher version numbers always mean lower fees
4. Pinning a version automatically applies security updates

Q1. You want consistent output format for inquiry classification. The prompt must stay short, and you have only one typical example. Which technique fits best?

1. Prompt chaining (run several prompts in order)
2. Zero-shot prompt (instructions only)
3. Few-shot prompt (several examples)
4. One-shot prompt (show one example)

Q2. Which task does Chain-of-Thought help most?

1. Brightening image colors
2. Writing a one-word greeting
3. Translating a single word
4. Problems requiring step-by-step reasoning

Q1. Which correctly describes how RAG (retrieval-augmented generation) works?

1. Registering model answers in a search engine for other users to see
2. Converting the question to an image and returning image search results as is
3. Retrieving relevant information and adding it to the prompt to generate the answer
4. Training the model on all answers in advance and doing no search at all

Q2. When is RAG most appropriately preferred over fine-tuning?

1. You want to fix the model's tone to a specific character's way of speaking
2. You want to raise accuracy on a specialized classification task with many examples
3. You want the model to deeply internalize a specific output format
4. You want answers with evidence from frequently updated internal information

Q1. A RAG chat over internal documents answered based on an abolished old rule. What is the most appropriate explanation of the cause?

1. Answer accuracy depends on model size; the quality of retrieval results has no effect
2. The content and freshness of retrieved documents determine the accuracy of generated answers
3. With a high temperature, retrieval results are ignored and answers come from pretrained knowledge
4. In RAG the model favors pretrained knowledge and uses retrieved documents only as a supplement

Q2. For product descriptions, you want to avoid unlikely, unnatural words while keeping some variety of expression. Which setting should you adjust?

1. Tighten safety settings to block inappropriate words
2. Lower top-p so the next word is chosen from candidates with high cumulative probability
3. Lower the output token limit to shorten text and reduce unnatural words
4. Raise temperature so the next word is chosen from more varied candidates