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

Amazon Bedrock introduces reinforcement fine-tuning, enabling developers to build accurate AI models via feedback without ML expertise.

• •Delivers 66% accuracy gains over base models using reward functions instead of large labeled datasets.
• •Supports RLVR for objective tasks (code, math) and RLAIF for subjective tasks (instruction following).
• •Works with Amazon Nova 2 Lite; training data from invocation logs or JSONL files in S3.
• •Reward functions use custom Python on AWS Lambda or foundation models as evaluators.