PRODUCTION SCENARIO
A retailer runs a returns classifier on a large Amazon Bedrock model. Accuracy is excellent, but cost and latency are too high for peak season. The team has thousands of production prompts and very few labeled answers.
Which approach will meet these requirements?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Model distillation
Distillation transfers knowledge from a larger, more capable teacher model to a smaller, faster, more cost-efficient student. Amazon Bedrock generates the teacher's responses for the customer's own prompts and uses them to fine-tune the student, so a large labeled dataset is not required. Supervised fine-tuning would need the labels the team lacks.