IBM C1000-185 試験概要:

IBM C1000-185 試験シラバストピック:

IBM watsonx Generative AI Engineer - Associate 認定 C1000-185 試験問題：

問題 #1
You're designing a prompt that should generate text for product descriptions using a generative AI model. You want to limit the model's word choices to ensure coherence while still allowing some flexibility for creativity. You decide to use Top-P (nucleus) sampling to achieve this.
Which of the following settings for the Top-P parameter is most appropriate to strike a balance between creativity and coherence?

A. Top-P = 0.3
B. Top-P = 1.0
C. Top-P = 0.1
D. Top-P = 0.9

問題 #2
In the context of Generative AI models, which scenario best illustrates the use of greedy decoding for determining the output sequence?

A. Selecting the token with the highest probability at each step, without considering future possibilities
B. Selecting the token that maximizes the cumulative probability of the sequence, considering future tokens
C. Randomly selecting tokens from the top-k highest probability tokens at each step
D. Averaging the probabilities of multiple tokens at each step to generate a sequence

問題 #3
You are fine-tuning a general-purpose language model on a medical dataset to generate summaries of patient consultations. After fine-tuning, you notice that the model sometimes generates hallucinations-statements that are factually incorrect or irrelevant to the specific domain. You suspect that the fine-tuning process did not sufficiently align the model with the medical domain.
Which of the following is the most effective technique to reduce hallucinations during fine-tuning?

A. Add more general-purpose data to the fine-tuning dataset
B. Use domain-specific tokenization during fine-tuning
C. Increase the model's batch size during training
D. Increase the number of layers in the model

問題 #4
You are tasked with fine-tuning a generative AI model for text data using synthetic data created through the IBM watsonx platform's user interface. The data you are working with is skewed, containing mostly outliers, and you need to ensure that the synthetic data mimics the distribution accurately.
Which algorithm would be most appropriate for generating synthetic data that mirrors the original distribution, considering the Anderson-Darling test for normality?

A. Generative Adversarial Networks (GANs)
B. Decision Trees
C. Bootstrapping
D. Anderson-Darling Based Synthetic Data Generation (ADS-DG)

問題 #5
You are designing prompt templates for an automated content generation tool that your team uses for marketing copy. The tool must balance cost efficiency with the generation of high-quality outputs. The prompts should be structured to avoid excessive token usage while ensuring that the model provides coherent and actionable content.
Which of the following techniques would best optimize the design of cost-effective prompt templates?

A. Designing prompts that incorporate redundancy to ensure model outputs address all possible user needs.
B. Using placeholders in the prompt templates for dynamic insertion of variables specific to each task.
C. Embedding explicit stop words within the prompt to control the length of generated outputs.
D. Creating templates that provide extensive background information to give the model context.

解説：