Introducing Nova Forge SDK, a seamless way to customize Nova models for enterprise AI
New SDK reduces LLM customization complexity by handling infrastructure, recipes, and dependency management automatically.
Amazon has introduced the Nova Forge SDK, a comprehensive developer toolkit designed to streamline the customization of its Nova large language models (LLMs) for enterprise applications. The SDK addresses a critical pain point: the technical complexity and high barrier to entry associated with fine-tuning general-purpose models for specific business domains. It provides a unified interface that spans the entire customization lifecycle—from data preparation and training job management to model deployment—across Amazon's AI platforms, including Bedrock and SageMaker AI. By automating undifferentiated heavy lifting like dependency management, image selection, and recipe configuration, the SDK empowers development teams to focus on experimentation and iteration rather than infrastructure.
The SDK is architected in three core layers to simplify the workflow. The Input Layer allows developers to specify parameters like hardware, platform, IAM roles, training methods, and hyperparameters. The Customizer Layer then takes these inputs and automatically builds the appropriate recipe configurations to launch the training job. Finally, the Output Layer delivers the results, including model artifacts, CloudWatch logs, and ML Flow metrics, ready for deployment on SageMaker AI or Bedrock. A key feature is its support for starting customization from early model checkpoints and blending proprietary datasets with Amazon's curated data, which helps combat catastrophic forgetting—the phenomenon where models lose base capabilities during fine-tuning. This approach gives enterprises both streamlined workflows for common tasks and the flexibility to access underlying service SDKs for advanced use cases.
- Unified toolkit supporting customization across Amazon Bedrock and SageMaker AI, including methods like SFT, DPO, and RFT.
- Automates infrastructure and recipe configuration to lower the technical barrier, handling dependency management and image selection.
- Architected in three layers (Input, Customizer, Output) to streamline the workflow from data prep to model deployment.
Why It Matters
Enables enterprises to efficiently build specialized AI models on proprietary data, accelerating domain-specific AI adoption without requiring deep machine learning expertise.