PRODUCTION SCENARIO
A financial technology company wants a model to produce SQL that runs correctly against its data warehouse. The company has 4,000 natural-language questions and a test harness that executes generated SQL and scores the result. It has no gold-standard SQL for those questions.
Which Amazon Bedrock customization method will meet these requirements?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Reinforcement fine-tuning
Reinforcement fine-tuning learns from feedback scores instead of labeled input-output pairs. Reward functions defined with AWS Lambda evaluate the quality of each response, which is exactly what the execution harness already provides. Supervised fine-tuning and distillation both need target outputs that the company does not have.