C1000-185 無料問題集「IBM watsonx Generative AI Engineer - Associate」

（A）Decreasing the number of tokens allowed in the output to encourage more creative responses.

（B）Increasing the temperature parameter to make the model more deterministic and consistent in its responses.

（C）Altering the tone of the prompt to make it more aligned with the brand's voice and customer expectations.

（D）Modifying the stop sequence to ensure the model responds with shorter answers for concise customer interactions.

（A）Setting the maximum tokens to 50 and minimum tokens to 150

（B）Setting the minimum tokens to 150 and maximum tokens to 50

（C）Setting the minimum tokens to 50 and maximum tokens to 150

（D）Setting no minimum token value but configuring the maximum tokens to 150

（A）Soft prompts offer higher flexibility by dynamically adjusting the model's behavior based on pre-learned embeddings rather than relying on static instructions.

（B）Soft prompts allow for more human oversight by embedding specific instructions into the model's generation process.

（C）Soft prompts are preferable in scenarios where the task requires rigid structure and adherence to specific predefined guidelines.

（D）Soft prompts produce more deterministic outputs because they are optimized for a specific task through manual input adjustment.

（A）When real-time similarity search over high-dimensional embeddings is needed for large-scale unstructured text data

（B）When the data consists primarily of binary files such as images and videos, and full-text search is required

（C）When the corpus consists mainly of short, structured text like JSON records and traditional SQL indexing will suffice

（D）When the text corpus consists entirely of predefined categories that can be handled by simple keyword matching algorithms

（A）A soft prompt is a fixed string of text used in fine-tuning, while a hard prompt adjusts dynamically based on input data.

（B）Soft prompts are more readable and natural, whereas hard prompts consist of short, technical instructions.

（C）Hard prompts are less efficient because they need to be re-trained with each task, while soft prompts are more versatile and adaptive across multiple tasks.

（D）A hard prompt explicitly specifies all constraints, while a soft prompt relies on implicit learning from continuous inputs during training.

（A）Few-shot prompting is used when the model is trained on supervised learning, while zero-shot prompting works only with unsupervised models.

（B）Zero-shot prompting requires no examples in the prompt, while few-shot prompting provides the model with one or more examples to clarify the task.

（C）Zero-shot prompting requires retraining the model with additional data, while few-shot prompting uses a pre-trained model without retraining.

（D）Zero-shot prompting provides the model with examples, while few-shot prompting does not.

（A）To return relevant documents or passages from a knowledge base by calculating semantic similarity between the query and documents.

（B）To generate new responses based on training data, without relying on external data sources.

（C）To perform pre-defined template matching on queries to retrieve documents that exactly match pre-configured templates.

（D）To perform entity recognition and classify documents based on specific keywords, ignoring the overall document meaning.

（A）Bootstrapping Algorithm

（B）K-Nearest Neighbors (KNN)

（C）Anderson-Darling Based Synthetic Data Generation (ADS-DG)

（D）Gaussian Mixture Models (GMMs)

（A）Boolean retriever, which uses logical operators (AND, OR, NOT) to filter documents based on keyword presence.

（B）Exact match retriever, which retrieves documents that exactly match the query string.

（C）Dense retriever, which converts both the query and documents into vectors and retrieves based on similarity in the vector space.

（D）Hierarchical retriever, which first applies keyword-based filters and then ranks documents using machine learning models.

（A）Embed the prompt version directly into the API request body so that the deployed model can select the correct prompt dynamically at runtime.

（B）Store all prompt versions directly in the model's code repository, updating the main branch with each new version.

（C）Keep prompt versions in an external document management system and manually track which versions are deployed in the application.

（D）Use a deployment space in IBM watsonx to version your prompts, assigning unique tags to each version and ensuring rollback capabilities.

（A）The model's size should be reduced to fit within the memory constraints of the deployment environment, even if it leads to a loss of precision in its outputs.

（B）You should enable quantization of the model to optimize inference performance without compromising accuracy in a production environment.

（C）The model should be fine-tuned with domain-specific data before deployment, as Watsonx requires fine-tuning for all foundation models before they can be deployed.

（D）You must ensure that the model is compatible with IBM's foundation model API, as Watsonx only allows deployment of models that integrate with its foundation model architecture.

（A）Enable auto-validation in the UI to automatically discard any generated data that doesn't align perfectly with training data patterns.

（B）Only use synthetic data that matches predefined patterns exactly to avoid any incoherence or errors.

（C）Generate as much data as possible, then use external validation tools to filter out problematic responses.

（D）Manually adjust the temperature parameter in the UI to lower values to reduce randomness and increase coherence in the generated text.

（A）Store the pre-trained generative model's parameters in SingleStore and use LangChain to retrieve embeddings from the database for training the model.

（B）Use LangChain to create an LLM chain that integrates with SingleStore for embedding retrieval, where the retrieved documents are used as context for generating responses.

（C）Use LangChain to generate embeddings and directly store the generated responses in SingleStore without any retrieval mechanism.

（D）Use LangChain to fine-tune the generative model, then store the embeddings in SingleStore, and use SQL queries to retrieve documents based on exact keyword matches.