Task Statement 3.3: Describe the Training and Fine-Tuning Process for FMs. AIF-C01 Practice Questions (Page 3)

Questions 11–15

1. 11foundation · easy

   What is a key data governance consideration when fine-tuning a foundation model?

   • AEnsuring the data complies with relevant regulations and privacy laws
   • BStoring all data in a single location for simplicity
   • CUsing only unlabeled data to avoid privacy issues
   • DMaximizing the size of the dataset regardless of quality
   Reveal Solution Discussion Select an answer first
2. 12expert · hard

   A government agency is fine-tuning a foundation model to summarize public policy documents. They have a large dataset of documents, but it is imbalanced: 80% of the documents are from the last 5 years, and 20% are from older decades. The older documents use different terminology and formatting. The agency needs the model to summarize both old and new documents accurately. What is the MOST effective data curation strategy?

   • AIncrease the overall dataset size by adding more recent documents.
   • BUse RLHF to have policy experts rank summaries of older documents.
   • CRemove all older documents to focus on the most recent and relevant policy language.
   • DBalance the dataset by including more older documents or upsampling them so the model sees a representative mix of both eras.
   Reveal Solution Discussion Select an answer first
4. 13expert · hard

   A chatbot company has fine-tuned a foundation model using instruction tuning to answer customer queries. The model is technically accurate but sometimes produces responses that are overly verbose, and it occasionally uses offensive language when users are frustrated. The company has a dataset of human preferences ranking responses from helpful to harmful. They also have a strict policy that the model must never use offensive language. Which approach BEST addresses both the verbosity and the safety concern?

   • AUse instruction tuning with additional examples of concise and safe responses.
   • BUse continuous pre-training on a dataset of polite conversations.
   • CUse knowledge distillation from a smaller, safer model.
   • DUse RLHF with a reward model trained on the preference data, and include a strong penalty for offensive language in the reward function.
   Reveal Solution Discussion Select an answer first
5. 14application · medium

   A legal technology company has a general-purpose foundation model that performs well on standard text tasks but struggles with legal terminology, case citations, and the specific writing style of court documents. They have a large corpus of unlabeled court records and legal opinions. Which approach would MOST effectively improve the model's performance on legal text while preserving its general language abilities?

   • AFine-tune the model directly on the unlabeled legal corpus using a masked language modeling objective.
   • BDistill the general-purpose model into a smaller model trained only on legal text.
   • CPerform continuous pre-training on the unlabeled legal corpus, then optionally fine-tune on a small labeled set for the final task.
   • DUse RLHF with legal experts rating the model's outputs on legal queries.
   Reveal Solution Discussion Select an answer first
6. 15foundation · easy

   What is a primary benefit of knowledge distillation?

   • AIt eliminates the need for labeled data
   • BIt makes the model better at following instructions
   • CIt increases the model's accuracy on all tasks
   • DIt reduces the model's size and inference cost while maintaining much of its performance
   Reveal Solution Discussion Select an answer first