Amazon Bedrock adds reinforcement fine-tuning simplifying how developers build smarter, more accurate AI models

This article announces reinforcement fine-tuning in Amazon Bedrock, a new model customization capability that enables developers to build smarter, more accurate AI models using feedback-driven training without requiring deep ML expertise.

• Reinforcement fine-tuning delivers 66% accuracy gains on average over base models
• Uses reward functions to evaluate responses instead of requiring large labeled datasets
• Supports two approaches: RLVR for objective tasks and RLAIF for subjective tasks
• Currently works with Amazon Nova 2 Lite, with additional models coming soon
• Automates workflow with training data from API logs or uploaded datasets
• Includes seven ready-to-use reward function templates for common use cases
• Data remains secure within AWS environment with VPC and KMS encryption support
• Models can be deployed and tested in Amazon Bedrock playground with one click

Amazon Bedrock's reinforcement fine-tuning simplifies advanced model customization, making it accessible to developers without specialized ML expertise or complex infrastructure setup.