IBM watsonx Generative AI Engineer - Associate : C1000-185

In a Retrieval-Augmented Generation (RAG) system, embeddings play a central role in linking input queries with relevant external knowledge. Different embedding models can be used to generate these embeddings.
Which of the following embedding models is best suited for capturing semantic meaning in text for use in a RAG system?

• A. One-Hot Encoding
• B. Word2Vec
• C. Latent Dirichlet Allocation (LDA)
• D. Bag-of-Words (BoW)

You are tasked with managing multiple versions of a generative AI model in a production environment. The team has decided to utilize IBM watsonx deployment spaces to version control both models and prompts.
What is the most efficient way to organize these assets within the deployment space?

• A. Store all versions of both models and prompts in a single deployment space, and use comments to track differences between versions.
• B. Assign version tags to both models and prompts within a single deployment space, using metadata and associations to track prompt-model pairings for each version.
• C. Use a third-party version control tool like Git to manage prompt and model versions, syncing with the deployment space only for final deployment.
• D. Create separate deployment spaces for each model and prompt version, linking the spaces manually to ensure consistency between model and prompt versions.

As an IBM Watsonx Generative AI engineer, you are tasked with creating a chatbot for a public-facing service. One key concern is ensuring that the model does not generate or propagate hate speech, abusive content, or profanity. To mitigate these risks, you must implement appropriate controls.
Which of the following is the best approach to mitigate hate speech, abuse, and profanity from being generated by your AI model?

• A. Fine-tune the model with a highly curated dataset that contains labeled examples of hate speech, abuse, and profanity for the model to learn to avoid.
• B. Apply a simple word-level blacklist filter to detect and remove harmful content from the model output.
• C. Train the model only on data that excludes all user-generated content to prevent exposure to harmful language.
• D. Use IBM Watsonx's HAP (Hate, Abuse, and Profanity) filter to dynamically detect and block harmful content at inference time.

You are optimizing the generative AI model in IBM watsonx to balance creativity and coherence. You want to use a decoding method that dynamically adjusts the token probability threshold based on cumulative probabilities, thus ensuring the model generates coherent outputs while still allowing for some creativity.
Which parameter should you adjust, and what is the optimal setting?

• A. Set top-p to 0.9 and temperature to 0.7
• B. Set temperature to 1 and top-p to 0.2
• C. Set temperature to 1.5 and top-p to 0.9
• D. Set temperature to 0 and top-p to 0.5

You are tasked with optimizing a prompt-tuned large language model (LLM) using IBM Watsonx for a customer service chatbot. The chatbot needs to handle a variety of tasks, such as answering frequently asked questions (FAQs), providing detailed product descriptions, and troubleshooting user issues.
What is the most appropriate task to focus on during the initial tuning experiment?

• A. Tune the model for text summarization, condensing user queries into shorter forms.
• B. Optimize the model for extractive question-answering from a predefined knowledge base.
• C. Fine-tune the model to generate product descriptions using longer contextual prompts.
• D. Focus on prompt-tuning the model for multi-turn dialogue to simulate more natural conversations.