AI-900 Practice Question: Describe features of generative AI workloads on Azure

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  A feature for adjusting model parameters in real time based on user feedback during deployment

  Why it's wrong here

  Real-time parameter adjustment based on user feedback during deployment is a form of dynamic configuration or online learning, not the offline fine-tuning process used in Azure OpenAI. Fine-tuning requires preparing a curated dataset and running a training job that updates the model's weights before the model is deployed. Feedback loops during deployment typically inform prompt engineering, A/B testing, or future retraining, but they do not constitute fine-tuning in the AI-900 sense.

• ✓

  Training a base model on domain-specific JSONL conversation examples to adapt its behaviour

  Why this is correct

  Fine-tuning in Azure OpenAI means taking a pre-trained base model (such as GPT-4o-mini or GPT-3.5-Turbo) and further training its weights on a custom dataset formatted as JSONL conversation examples, each containing system, user, and assistant role messages. This supervised training modifies the model's behavior to produce a consistent style, follow a specific output format, or incorporate domain-specific knowledge. The JSONL structure is essential because it teaches the model the expected dialogue flow and response patterns for your use case.

• ✗

  A no-code interface for adjusting temperature and top_p settings without writing code

  Why it's wrong here

  Adjusting temperature and top_p is an inference-time sampling configuration that controls the randomness and diversity of generated tokens, and it can be done through a no-code interface without retraining the model. Fine-tuning is fundamentally different: it is an actual training job that updates the neural network's parameters by processing custom examples, not just changing generation settings. These sampling parameters apply to any model, base or fine-tuned, and do not alter the model's learned knowledge or capabilities.

• ✗

  Restricting the model to only generate responses related to topics in your training data

  Why it's wrong here

  Restricting a model to generate only topically related responses is typically achieved with system prompts, few-shot examples, or content filters/moderation layers, not by fine-tuning. Fine-tuning adjusts the model's behavior and style through training data, but it cannot enforce a hard boundary on what topics the model will accept; the model may still produce out-of-domain content if prompted, and safety guardrails are needed. Confusing fine-tuning with topic blocking is a common misconception because fine-tuning allows domain specialization but does not 'lock' the model to that domain.

Go deeper

Related to this question