Task Statement 3.3: Describe the Training and Fine-Tuning Process for FMs. AIF-C01 Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   Which type of data is typically used during the pre-training of a foundation model?

   • ALarge, unlabeled text corpora from diverse sources
   • BSmall, labeled datasets specific to a single task
   • CCurated instruction-response pairs
   • DHuman preference rankings of model outputs
   Reveal Solution Discussion Select an answer first
2. 17application · medium

   A customer support platform is building a chatbot that must follow complex, multi-step instructions from users, such as 'Refund the order and then send a confirmation email to the customer.' The base foundation model understands language but often ignores parts of the instruction or responds with irrelevant information. The team has collected a dataset of example instructions paired with ideal assistant responses. Which training approach should they use to improve the model's ability to follow user prompts?

   • AInstruction tuning on the dataset of instruction-response pairs.
   • BKnowledge distillation from a larger model that already follows instructions well.
   • CContinuous pre-training on a large corpus of customer support tickets.
   • DRLHF using customer satisfaction ratings on the chatbot's responses.
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A marketing agency is fine-tuning a foundation model to generate ad copy for a new product line. They have two candidate datasets: Dataset A has 2,000 examples of ad copy, but 30% of the labels are incorrect. Dataset B has 500 examples, all accurately labeled and representative of the target audience. Which dataset is likely to produce a better fine-tuned model, and why?

   • ADataset B, because smaller datasets always produce better models.
   • BDataset B, because accurately labeled, representative data is more important than a larger dataset with noisy labels.
   • CDataset A, because the model can correct the label errors during fine-tuning.
   • DDataset A, because the larger size will help the model generalize better despite the label errors.
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   A social media company is fine-tuning a foundation model to generate short, engaging post summaries. After initial fine-tuning, the model produces grammatically correct summaries, but they are often bland and do not match the company's desired tone. The company has a team of editors who can rank different summaries from best to worst. Which technique should they use to align the model's output with human preferences?

   • AReinforcement learning from human feedback (RLHF), using the editors' rankings as a reward signal.
   • BKnowledge distillation from a larger, more engaging model.
   • CInstruction tuning on a dataset of generic summarization instructions.
   • DContinuous pre-training on a larger corpus of social media posts.
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   Why is accurate labeling important in a fine-tuning dataset?

   • AIt ensures the model learns the correct input-output mapping
   • BIt makes the model smaller
   • CIt reduces the need for a large dataset
   • DIt speeds up the pre-training phase
   Reveal Solution Discussion Select an answer first