Customizing models with Amazon SageMaker AI
Amazon SageMaker AI model customization is a capability that transforms the traditionally complex and time-consuming process of customizing AI models from a months-long endeavor into a streamlined workflow that can be completed in days. This feature addresses the critical challenge faced by AI developers who need to customize foundation models with proprietary data to create highly differentiated customer experiences. Detailed customization documentation, including step-by-step guides and advanced configuration options, is provided in this SageMaker AI guide. For a brief overview of Nova model customization, see Customize and fine-tune with SageMaker in the Amazon Nova User Guide.
The capability includes a new guided user interface that understands natural language requirements, with a comprehensive suite of advanced model customization techniques, all powered by serverless infrastructure that eliminates the operational overhead of managing compute resources. Whether you're building legal research applications, enhancing customer service chatbots, or developing domain-specific AI agents, this feature accelerates your path from proof-of-concept to production deployment.
Features in Model Customization powered by Amazon Bedrock Evaluations may securely transmit data across Amazon Web Services Regions within your geography for processing. For more information, access Amazon Bedrock Evaluations documentation.
Key concepts
Serverless training
A fully managed compute infrastructure that abstracts away all infrastructure complexity, allowing you to focus purely on model development. This includes automatic provisioning of GPU instances (P5, P4de, P4d, G5) based on model size and training requirements, pre-optimized training recipes that incorporate best practices for each customization technique, real-time monitoring with live metrics and logs accessible through the UI, and automatic cleanup of resources after training completion to optimize costs.
Model customization techniques
Comprehensive set of advanced methods including supervised fine-tuning (SFT), direct preference optimization (DPO), reinforcement learning with verifiable rewards (RLVR), and reinforcement learning with AI feedback (RLAIF).
Custom model
A specialized version of a base foundation model that has been adapted to a specific use case by training it on your own data, resulting in an AI model that retains the general capabilities of the original foundation model while adding domain-specific knowledge, terminology, style, or behavior tailored to your requirements.
AI model customization assets
Resources and artifacts used to train, refine, and evaluate custom models during the model customization process. These assets include datasets, which are collections of training examples (prompt-response pairs, domain-specific text, or labeled data) used to fine-tune a foundation model to learn specific behaviors, knowledge, or styles, and evaluators, which are mechanisms for assessing and improving model performance through either reward functions (code-based logic that scores model outputs based on specific criteria, used in RLVR training and custom scorer evaluation) or reward prompts (natural language instructions that guide an LLM to judge the quality of model responses, used in RLAIF training and LLM-as-a-judge evaluation).
Model package group
A collection container that tracks all logged models from training jobs, providing a centralized location for model versions and their lineage.
Logged model
The output created by SageMaker AI when running serverless training jobs. This can be a fine-tuned model (successful job), a checkpoint (failed job with checkpoint), or associated metadata (failed job without checkpoint).
Registered model
A logged model that has been marked for formal tracking and governance purposes, enabling full lineage and lifecycle management.
Lineage
The automatically captured relationships between training jobs, input datasets, output models, evaluation jobs, and deployments across SageMaker AI and Amazon Bedrock.
Cross-account sharing
The ability to share models, datasets, and evaluators across Amazon accounts using Amazon Resource Access Manager (RAM) while maintaining complete lineage visibility.