Domain 2: Fundamentals of Generative AI

Question 1

What does the term 'temperature' refer to when generating text with a transformer-based large language model?

A parameter that adjusts the learning rate during the model's pre-training and fine-tuning phases A metric that measures the computational heat generated during the model's inference process A setting that determines the maximum number of tokens the model can produce in a single response A parameter that controls the randomness and creativity of the model's output tokens

Correct Answer: A parameter that controls the randomness and creativity of the model's output tokens

Temperature is a hyperparameter used during inference that scales the logits (raw output scores) before the softmax function, thereby controlling how random or deterministic the model's token selection is. A higher temperature produces more diverse and creative outputs, while a lower temperature makes the model more focused and predictable.

Computational heat has nothing to do with the temperature parameter in language models. Adjusting the learning rate is a separate training hyperparameter unrelated to temperature. Maximum token length is controlled by a different parameter, often called max_tokens or max_length.

Key concept tested: understanding inference-time parameters that control LLM output behavior.

Question 2

A team is new to AI and wants to build a simple text-generation application. They learn that transformer-based large language models (LLMs) are trained on massive datasets before being used. What is the main benefit of using a pre-trained foundation model for their project?

It contains a curated small dataset, so the team benefits from highly focused and specialized language patterns It already has broad language knowledge, so the team can use it for tasks with minimal additional training It already has broad language knowledge, so the team must retrain it fully on their own dataset before use It has been optimized for one specific task, so the team can deploy it for their exact text-generation need

Correct Answer: It already has broad language knowledge, so the team can use it for tasks with minimal additional training

A pre-trained foundation model has already learned broad language patterns, grammar, and knowledge from massive datasets, enabling teams to leverage it for downstream tasks like text generation with minimal or no additional training through techniques such as prompting or light fine-tuning.

The option suggesting full retraining is wrong because it defeats the purpose of using a pre-trained model. The option about being optimized for one specific task is incorrect because foundation models are general-purpose by design, not task-specific. The option about a curated small dataset is wrong because foundation models are trained on massive, diverse datasets, not small focused ones.

Key concept tested: the value of transfer learning and pre-trained foundation models in reducing development effort.

Question 3

Which type of machine learning providers are included in the Amazon SageMaker JumpStart model hub?

Only AWS first-party models developed and trained by Amazon's internal research teams Exclusively open-source community models curated from GitHub and similar repositories Both proprietary and open-source model providers such as AI21 Labs, Hugging Face, and Meta Only proprietary commercial model providers that have enterprise licensing agreements with AWS

Correct Answer: Both proprietary and open-source model providers such as AI21 Labs, Hugging Face, and Meta

Amazon SageMaker JumpStart model hub provides access to a wide variety of foundation models from both proprietary and open-source providers, including AI21 Labs, Hugging Face, Meta, Stability AI, Cohere, and others, making it a diverse marketplace for pre-trained models.

The option stating only AWS first-party models is incorrect because JumpStart specifically highlights third-party model availability. The option limiting to open-source only is wrong because proprietary providers like AI21 Labs are also included. The option restricting to proprietary commercial providers only is incorrect because open-source models from providers like Hugging Face and Meta are prominently featured.

Key concept tested: understanding the breadth and diversity of model providers available in SageMaker JumpStart.

Domain 2: Fundamentals of Generative AI

AWS AIF-C01 - Domain 2: Fundamentals of Generative AI Example Questions

Question 1

Question 2

Question 3

Validate Your AI Knowledge on AWS