Task Statement 3.3: Describe the Training and Fine-Tuning Process for FMs. AIF-C01 Practice Questions (Page 1)

Questions 1–5

1. 1foundation · easy

   What is the purpose of continuous pre-training?

   • ATo train a model on domain-specific data to improve performance in that domain
   • BTo adapt a model to a specific task using labeled examples
   • CTo compress a large model into a smaller one
   • DTo align model outputs with human preferences
   Reveal Solution Discussion Select an answer first
2. 2expert · hard

   A company has a large, accurate foundation model for sentiment analysis that runs in the cloud. They want to deploy a version on edge devices with limited memory and compute. They have a small, labeled dataset of 1,000 examples for the target domain. The team is considering two approaches: (1) fine-tune the large model on the labeled data and then distill it into a small model, or (2) distill the large model first and then fine-tune the small model on the labeled data. Which approach is likely to yield a better small model, and why?

   • ABoth approaches will produce identical results because distillation and fine-tuning are independent.
   • BApproach 2, because the small model will learn the domain better from the labeled data without the large model's bias.
   • CApproach 1, because fine-tuning the large model first improves its domain accuracy, which the distilled small model can then learn from.
   • DApproach 2, because the small model will be easier to fine-tune after distillation.
   Reveal Solution Discussion Select an answer first
4. 3foundation · easy

   What is knowledge distillation in the context of foundation models?

   • ATransferring knowledge from a large model to a smaller model while preserving performance
   • BUsing human feedback to align a model's outputs with preferences
   • CTraining a large model on a distilled version of the dataset
   • DTraining a model on a small labeled dataset to adapt it to a specific task
   Reveal Solution Discussion Select an answer first
5. 4application · medium

   A news organization has a general-purpose foundation model that writes articles well but frequently uses outdated terminology and does not understand recent events. They have a large archive of recent news articles (unlabeled) and a small set of labeled examples of their preferred writing style. What is the MOST effective sequence of training steps to improve the model?

   • AFine-tuning on the labeled style examples, followed by continuous pre-training on the recent news articles.
   • BContinuous pre-training on the recent news articles, followed by fine-tuning on the labeled style examples.
   • CKnowledge distillation from a smaller model trained on recent news.
   • DRLHF using reader feedback on article quality.
   Reveal Solution Discussion Select an answer first
6. 5foundation · easy

   Which dataset is required for fine-tuning a foundation model?

   • AA large unlabeled corpus of general text
   • BA dataset of instruction-response pairs
   • CA small labeled dataset relevant to the target task
   • DA dataset of human preference comparisons
   Reveal Solution Discussion Select an answer first