Task Statement 3.3: Describe the Training and Fine-Tuning Process for FMs. AIF-C01 Practice Questions (Page 2)

Questions 6–10

1. 6expert · hard

   A startup is fine-tuning a foundation model to detect hate speech in multiple languages. They have a large dataset of English examples but only a few hundred examples for Spanish and French. The model performs well on English but poorly on Spanish and French. The team has a limited budget and cannot collect more data. Which strategy is MOST likely to improve performance on Spanish and French without collecting new data?

   • AUse the English data to fine-tune the model, then evaluate and fine-tune further on the small Spanish and French datasets.
   • BUse RLHF with human feedback from Spanish and French speakers.
   • CUse continuous pre-training on the English data to improve the model's general language understanding.
   • DRemove the Spanish and French data and focus only on English to improve overall accuracy.
   Reveal Solution Discussion Select an answer first
2. 7foundation · easy

   What is reinforcement learning from human feedback (RLHF)?

   • AA method that uses human feedback to align model outputs with human preferences
   • BA method that trains a model on unlabeled data to learn language patterns
   • CA method that transfers knowledge from a large model to a smaller one
   • DA method that trains a model on instruction-response pairs
   Reveal Solution Discussion Select an answer first
4. 8foundation · easy

   Why is it important for fine-tuning data to be representative of the target domain and user population?

   • ATo make the model easier to deploy
   • BTo avoid bias and improve generalization to real-world use
   • CTo reduce the size of the dataset
   • DTo ensure the model is trained faster
   Reveal Solution Discussion Select an answer first
5. 9expert · hard

   A company is building a virtual assistant that must follow user instructions precisely, such as 'Book a flight and send the itinerary to my email.' The base model follows simple instructions but fails on multi-step ones. The team has a dataset of 10,000 instruction-response pairs and a smaller dataset of 1,000 human preference rankings. They want to maximize instruction-following accuracy. What is the MOST effective training strategy?

   • AFirst perform instruction tuning on the 10,000 pairs, then apply RLHF using the 1,000 preference rankings.
   • BOnly use RLHF with the 1,000 preference rankings, skipping instruction tuning.
   • CUse continuous pre-training on the instruction-response pairs before fine-tuning.
   • DOnly use instruction tuning on the 10,000 pairs, ignoring the preference data.
   Reveal Solution Discussion Select an answer first
6. 10foundation · easy

   Why is data curation important for fine-tuning a foundation model?

   • AIt speeds up the pre-training phase
   • BIt reduces the size of the model
   • CIt ensures the model learns from high-quality, relevant examples
   • DIt eliminates the need for labeled data
   Reveal Solution Discussion Select an answer first