After completing a prompt-tuning experiment, you notice that the model's accuracy in generating relevant responses is high, but the fluency and grammatical correctness of the outputs seem to be suboptimal.
What statistical metric would most directly indicate this issue, and what action should you take to improve the output?

• A. Perplexity score; apply additional language model fine-tuning on grammatical correctness.
• B. BLEU score; improve prompt engineering to ensure that the model focuses on fluency.
• C. ROUGE score; adjust the token generation limit to ensure longer outputs.
• D. F1 score; increase the training dataset size to improve overall accuracy.

You are deploying an AI model to a production environment using watsonx.ai. The model is a fine-tuned GPT-based generative model designed to support real-time customer interactions. The deployment must ensure scalability, security, and maintain high availability.
Which deployment strategy should you choose to best meet these requirements?

• A. Deploy the model on an on-premises server with an isolated environment to ensure full control and security.
• B. Deploy the model on multiple virtual machines, manually balancing traffic between instances.
• C. Use a containerized deployment with Kubernetes on a multi-node GPU cluster to support auto-scaling and fault tolerance.
• D. Deploy the model on a single cloud-based GPU instance, and enable scheduled maintenance to minimize downtime.

When optimizing a prompt-tuned model, which parameter adjustment would most likely help prevent overfitting without negatively impacting the model's ability to generalize to unseen prompts?

• A. Reducing the model's learning rate
• B. Increasing the model's learning rate
• C. Increasing the model's dropout rate
• D. Removing weight decay entirely

You are training a chatbot to handle customer inquiries for a telecommunications company. To augment your training data, you decide to generate synthetic data using IBM's InstructLab platform. You aim to improve the model's ability to handle rare or edge-case scenarios, such as technical issues with specific device models. You are following the LAB (Large-scale Alignment for chatBots) methodology to ensure alignment of the chatbot's responses with company policies.
Which of the following steps is most aligned with LAB methodology principles for generating synthetic data in this case?

• A. Manually write every possible customer inquiry scenario to ensure quality
• B. Generate synthetic customer inquiries and responses based on company policy guidelines
• C. Use random generation to create a diverse range of responses without constraints
• D. Generate synthetic data without aligning it with company policies to maximize diversity

You are fine-tuning a generative model to generate text-based responses in a customer service chatbot. You want to ensure the responses are concise and relevant, without causing the model to produce overly long or irrelevant output.
Which of the following parameters and stopping criteria would be most effective for achieving this goal?

• A. Set a low top-k value and implement a repetition penalty with a low maximum token limit.
• B. Increase the temperature to 1.5 and set a high maximum token limit.
• C. Use beam search decoding with a low beam width and a high repetition penalty.
• D. Use greedy decoding with no repetition penalty and a high stopping probability threshold.