Exams / AIF-C01 / Training and fine-tuning / Question 1
AIF-C01 practice question: Training and fine-tuning (1 of 3)
Applications of Foundation Models A free question from our original AIF-C01 practice exam. Try it, then open the answer to see why each option is right or wrong.
A media group is comparing ways to adapt foundation models. Match each description to the technique it describes.
Match each item to one answer. An answer can be used more than once, and some may not be used.
Items
- Keep training an existing model on a large, unlabelled collection of industry documents
- Train an existing model on labelled prompt and response pairs for one task
- Train a smaller model on responses generated by a larger model for a use case
- Train a reward model from human rankings of responses, then optimise the model against it
- Put a few worked examples in the prompt, with no change to the model's weights
Answers
- AA. Continued pre-training
- BB. Distillation
- CC. In-context learning
- DD. Pre-training from scratch
- EE. RLHF
- FF. Supervised fine-tuning
Show the answer and explanation
Correct matches
- Keep training an existing model on a large, unlabelled collection of industry documents: A. Continued pre-training
- Train an existing model on labelled prompt and response pairs for one task: F. Supervised fine-tuning
- Train a smaller model on responses generated by a larger model for a use case: B. Distillation
- Train a reward model from human rankings of responses, then optimise the model against it: E. RLHF
- Put a few worked examples in the prompt, with no change to the model's weights: C. In-context learning
About each answer
- A. Continues training an existing model on unlabelled data to build domain knowledge. It matches the first description.
- B. Transfers knowledge from a larger teacher model to a smaller, cheaper student model. It matches the third description.
- C. Guides the model with examples in the prompt, with no training. It matches the fifth description.
- D. Distractor: builds a brand new model from a huge dataset. Every description starts from an existing model.
- E. Reinforcement learning from human feedback uses human rankings to train a reward model. It matches the fourth description.
- F. Trains on labelled input and output pairs for a specific task. It matches the second description.
Explanation
Techniques differ in what data they need and what they cost. In-context learning needs no training. Fine-tuning needs labelled pairs. Continued pre-training needs large amounts of unlabelled domain text. Distillation uses a teacher model's responses to train a smaller student. RLHF uses human preferences to align the model. Pre-training from scratch is by far the most expensive.
Study
Keep practising
Answer questions with a score at the end, or practise just this topic with the explanation after each one. No sign-up needed.