You can customize the Amazon Nova models using the fine-tuning method with labeled proprietary data on Amazon Bedrock to gain more performance for your use case than the models provide out-of-the-box. That is, fine-tuning provides enhancements beyond what is gained with zero- or few-shot invocation and other prompt engineering techniques. You can fine-tune Amazon Nova models when a sufficient amount of high-quality, labeled training data that is available for the following use cases:
-
When you have a niche or specialized tasks in a specific domain.
-
When you want model outputs aligned with brand tone, company policies, or proprietary workflows.
-
When you need better results across a wide number of tasks, and thus need to introduce examples in training. This situation is in contrast to providing instructions and examples in prompts, which also impacts token cost and request latency.
-
When you have tight latency requirements and can benefit from smaller models that are tailored to a specific use case.
Topics
Available models
Fine-tuning is available for the following Amazon Nova models and their supported text, image, and video modalities.
-
Amazon Nova Micro
-
Amazon Nova Lite
-
Amazon Nova Pro
-
Amazon Nova Canvas
Performing custom fine-tuning
To perform custom fine-tuning with Amazon Nova models, you do the following:
-
Create a training dataset and a validation dataset (if applicable) for your customization task. For more information about preparing data, see the following:
-
If you plan to use a new custom IAM role, follow the instructions in Create a service role for model customization to create an IAM role with access to your data in Amazon S3 buckets. Or you can use an existing role or let the console automatically create a role with the proper permissions.
-
(Optional) Configure Encryption of Amazon Nova model customization jobs and artifacts, VPC, or both, for extra security.
-
Create a Fine-tuning job, controlling the training process by adjusting the hyperparameter values.
-
Analyze the results by looking at the training or validation metrics or by using model evaluation.
-
Purchase Provisioned Throughput for your newly created custom model.
-
Use your custom model as you would a base model in Amazon Bedrock tasks, such as model inference.
