AWS Introduces Customization for Amazon Nova Models with Ready-to-Use SageMaker Recipes

AWS is announcing a new set of customization capabilities for Amazon Nova—AWS’ family of understanding models—within Amazon SageMaker AI. Now featuring customization techniques such as pre-training, supervised fine-tuning, alignment, continued pre-training, and knowledge distillation, customers can configure Nova Micro, Nova Lite, and Nova Pro to best suit their desired outcomes.

This release recognizes the need for AWS customers powering a diverse range of generative AI use cases, many require model customizations to meet certain accuracy, cost, and latency demands.

With new customization techniques available as ready-to-use Amazon SageMaker recipes, customers can select and combine multiple techniques to meet their unique needs. These new techniques aim to augment Amazon Nova with the ability to effectively represent proprietary knowledge, workflows, and brand requirements with customers’ preferred mix of performance, cost, and flexibility, according to AWS.

Diving deeper, the new customizations include:

• Supervised fine-tuning (SFT) customizes model parameters using a training dataset of input-output pairs specific to your target tasks and domains, including both parameter-efficient fine-tuning (PEFT) and full fine-tuning (FFT)
• Alignment designed to guide model outputs toward desired preferences, including behavior that aligns with company brand or customer experience requirements, including direct preference optimization (DPO) and proximal policy optimization (PPO)
• Continued pre-training (CPT) broadens foundational model knowledge via self-supervised learning on large volumes of unlabeled proprietary data, including internal documents, transcripts, and business-specific content
• Knowledge distillation allows a larger “teacher” model to transfer knowledge to a smaller, faster, more cost-efficient “student” model, ideal for customers lacking in reference input-output samples

To learn more about AWS’ customization techniques for Amazon Nova, please visit https://aws.amazon.com/?nc2=h_home.