Customizing Amazon Nova 2.0 models
Amazon Nova offers the most comprehensive suite of customization options to adapt the foundation models to your specific business needs, from simple fine-tuning in Amazon Bedrock to advanced training pipelines in Amazon SageMaker.
Topics
Customization overview
Model customization allows you to specialize Amazon Nova models for your domain, use cases and quality requirements. You can choose from several customization techniques and platforms based on your technical requirements, data availability and desired outcomes.
Customization techniques:
-
Continued Pre-Training (CPT) - Teach models domain-specific knowledge using raw text data
-
Supervised Fine-Tuning (SFT) - Customize through input-output examples
-
Reinforcement Fine-Tuning (RFT) - Optimize using reward signals and human feedback
-
Distillation - Transfer knowledge from larger to smaller models
Customization platform availability
The following table summarizes the availability of customizations per platform.
Platform |
CPT |
SFT |
RFT |
Distillation |
|---|---|---|---|---|
Amazon Bedrock |
No |
Yes |
Yes |
Yes |
SageMaker AI Training Jobs |
No |
Yes |
Yes |
Yes |
SageMaker AI HyperPod |
Yes |
Yes |
Yes |
Yes |
Customization on Amazon Bedrock
Amazon Bedrock provides a fully managed fine-tuning experience for Amazon Nova models, making it easy to customize models without managing infrastructure.
Supported methods:
- Supervised Fine-Tuning (SFT)
-
Teach models through input-output examples to customize response style, format and task-specific behavior.
- Reinforcement Fine-Tuning (RFT)
-
Maximize accuracy and align the model with real-world feedback and simulations using reward signals.
- Model Distillation
-
Transfers knowledge from larger "teacher" models to smaller "student" models. This process creates efficient models that maintain a significant portion of the original model's performance. The teacher model generates responses to diverse prompts and these outputs train the student model to produce similar results. This approach is more effective than standard fine-tuning when you lack sufficient high-quality labeled data.
Note
For implementation details, see Customization on Amazon Bedrock.
Key features:
-
Fully managed infrastructure with no setup required
-
Simple API-based training job submission
-
Direct deployment to Amazon Bedrock inference endpoints
When to use Amazon Bedrock fine-tuning:
-
You need quick customization with minimal setup
-
Your use case fits standard fine-tuning patterns
-
You prefer flexible customization from simple to increasingly complex training
-
You want seamless integration with Amazon Bedrock inference
Customization on SageMaker AI
SageMaker AI provides advanced training capabilities when you need full control over the customization process, access to multiple training methods and the ability to build simple to increasingly complex training pipelines.
Available training methods:
- Continued Pre-Training (CPT)
-
Teaches models domain-specific knowledge at scale using raw text data. Ideal for specialized technical fields, legal documents, medical literature, or any domain with unique terminology and concepts. Requires large volumes of unlabeled text (billions of tokens recommended).
- Supervised Fine-Tuning (SFT)
-
Customizes models through direct input-output examples. Best for teaching specific response styles, formats and task behaviors. Supports text, image and video inputs. You can choose between Parameter Efficient Fine-Truning (PEFT) of Full Fine-Tuning methods. The prior is suited for adapter-based customization with 2k-10k samples. The latter is suited for full model customization with tens of thousands to hundreds of thousands of samples.
- Reinforcement Fine-Tuning (RFT)
-
Optimizes models using reward signals for complex problem-solving tasks like domain-specific reasoning, code generation, and scientific analysis. Best used after SFT establishes baseline capabilities. You can use regex patterns and custom python code for verifiable rewards, or LLM-as-a-Judge for non-verifiable rewards. You can also integrate any other custom reward functions through a Lambda function such as single-turn remote reward functions in a customer's own environment.
- Model Distillation
-
Transfers knowledge from larger "teacher" models to smaller "student" models. This process creates efficient models that maintain a significant portion of the original model's performance. The teacher model generates responses to diverse prompts and these outputs train the student model to produce similar results. This approach is more effective than standard fine-tuning when you lack sufficient high-quality labeled data.
Note
For implementation details on distillation, see Model distillation.
Advanced capabilities:
-
Iterative training - Chain multiple training methods (for example, SFT to RFT) with checkpoint reuse for targeted improvements
-
Reasoning Mode Support - Train Nova 2 models with explicit reasoning steps for complex analytical tasks
Infrastructure options:
- SageMaker AI Training Jobs
-
Managed training with automatic resource provisioning for streamlined model customization workflows.
- SageMaker AI HyperPod
-
Resilient, large-scale training clusters for enterprise workloads requiring maximum control and scale.