AI-900 Which Command Should the Administrator Use Practice Questions

Explanation: Azure OpenAI Service includes a content filtering system that detects and blocks harmful categories of content such as hate, violence, sexual, and self-harm. Fine-tuning adapts a model to a specific task but does not guarantee blocking undesired outputs. Prompt engineering can reduce harmful outputs by careful phrasing, but is not a reliable safety mechanism alone. Token limits restrict the length of output, not the nature of the content.

Explanation: Content filtering in Azure OpenAI allows you to filter out harmful content based on categories like hate, violence, and self-harm. Prompt engineering and few-shot learning guide the model's output but do not guarantee prevention of harmful content. Fine-tuning customizes the model on specific data but does not inherently filter harmful outputs. Therefore, content filtering is the most direct feature for blocking unwanted content.

Explanation: Azure OpenAI Service provides content filtering capabilities that automatically detect and block harmful or inappropriate content in both prompts and completions. For image generation, the DALL-E model is used, and content filtering is applied to the generated images. By configuring content filters, organizations can enforce safety guidelines and prevent the generation of offensive or unsafe visuals. GPT-4 and the Embeddings model are designed for text, not image generation. Fine-tuning customizes model behavior on training data but does not add content filtering.

Explanation: The Chat Completions API accepts a conversation history as a list of messages and generates context-aware responses. The Completions API is for single-turn prompts without history. The Embeddings API creates vector representations. Fine-tuning is for customizing a model on specific data.

Explanation: For imbalanced datasets, AUC_weighted is the recommended primary metric in Azure Automated Machine Learning because it calculates the area under the ROC curve weighted by class prevalence, making it robust to skew. Accuracy would be misleading (a model predicting all negative gets 95% accuracy), F1_score can be used but is not as robust, and log_loss focuses on prediction uncertainty.