MLA-C01 Practice Question 35: Model Distillation Bedrock
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 4 · question 5 of 10
A support team uses a large foundation model on Amazon Bedrock to classify and route incoming tickets. Accuracy is excellent, but at the current volume the model is too expensive and too slow. A smaller model in the same family is cheap and fast enough, but its accuracy on this task is noticeably worse. The team has 20,000 historical ticket prompts but no human-written responses or labels for them, and it has no budget to hire annotators.
Which Amazon Bedrock approach best closes the accuracy gap on the smaller model?
Show answer and explanation
Answer:
Model distillation transfers knowledge from a larger, more accurate *teacher* model to a smaller, faster, cheaper *student* model. In Amazon Bedrock the training data can be just a set of prompts. Bedrock generates responses from the teacher and uses them to fine-tune the student. That fits a team that has prompts but no labels, and it produces a model with the smaller model's cost and speed plus much of the larger model's accuracy on this task (option A).
- B is wrong because supervised fine-tuning needs *labeled* examples, meaning prompts paired with the desired outputs. Prompts alone give the model nothing to learn to produce, and the team has no budget to write those outputs. - C is wrong because Provisioned Throughput reserves capacity for the large model. It can steady latency, but it does not make the large model cheaper, and cost is half of the problem. - D is wrong because retrieval adds context to each prompt, but the retrieved tickets have no correct routing attached, so the smaller model gets no better at the task. It also makes every request longer and slower.
Reference: Customize a model with distillation in Amazon Bedrock