Amazon Nova 2.0 advanced fine-tuning capabilities
Amazon Nova 2.0 introduces enhanced fine-tuning capabilities on Amazon Bedrock, including supervised fine-tuning with reasoning content and reinforcement fine-tuning with reward-based optimization.
Supervised fine-tuning on Amazon Nova 2.0
Amazon Nova 2.0 supervised fine-tuning uses the same Converse API format as Amazon Nova 1.0 with optional reasoning content fields, allowing you to train models that show their thinking process before generating final answers.
Key features
Support for text, image and video inputs in user content blocks
Optional reasoning content in assistant responses to capture intermediate thinking steps
Homogeneous dataset requirements (choose text-only, text+image, or text+video)
Support for PNG, JPEG and GIF images
Support for MOV, MKV and MP4 videos
Configurable reasoning modes for training optimization
For detailed information about data preparation, format specifications and reasoning modes, see Supervised fine-tuning on Amazon Nova 2.0 in the Amazon Bedrock user guide.
Reinforcement fine-tuning (RFT)
Reinforcement fine-tuning optimizes Amazon Nova models using measurable feedback signals rather than exact correct answers. This approach excels when you can reliably measure response quality but defining exact correct outputs is challenging.
Key features
Reward-based optimization using custom reward functions
Support for subjective or multifaceted quality assessment
Ideal for mathematical problem-solving and code generation
Effective for scientific reasoning and structured data analysis
Programmatic success verification through execution results
OpenAI Reinforcement Fine-Tuning format with messages and reference answers
Evaluation-first approach to validate reward functions before scaling
For detailed information about RFT data format, dataset recommendations and training characteristics, see Reinforcement fine-tuning (RFT) for Amazon Nova models in the Amazon Bedrock user guide.
