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, and it needs the harness score to be the training signal.
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 produces. Supervised fine-tuning needs target outputs the company does not have, distillation trains the student on responses the teacher model generates rather than on the harness score, and continued pre-training adapts a model to a domain corpus rather than to a pass or fail signal.