Fine-tune Amazon Nova models for accurate email data extraction

Fine-tuning Amazon Nova models using Amazon SageMaker AI can achieve up to 94.77% extraction accuracy for email data extraction, reducing costs by 50% and hallucinations. Parcel Perform, an ecommerce business, collaborated with the AWS Generative AI Innovation Center to optimize Nova models, resulting in improved accuracy, latency, and cost. The solution uses supervised fine-tuning with Parameter-Efficient Fine-Tuning (PEFT) through Low-Rank Adaptation (LoRA) to customize models effectively with limited training data. This approach enables flexible deployment options, including on-demand inference and provisioned throughput on Amazon Bedrock.

⚡ Key Takeaways

• The fine-tuned Nova Micro model achieved up to 94.77% extraction accuracy, a 16.6 percentage point improvement over the baseline.
• Parameter-Efficient Fine-Tuning (PEFT) with Low-Rank Adaptation (LoRA) enables effective model customization with limited training data.
• Amazon Nova recipes, YAML configuration files, provide details for model customization jobs, including base model name, training hyperparameters, and optimization settings.
• The solution supports deployment through on-demand inference, provisioned throughput on Amazon Bedrock, or a SageMaker AI endpoint.
• Fine-tuning reduced inference latency by more than 30% and halved costs compared to the previous model.

Fine-tuning Amazon Nova models can significantly improve the accuracy and efficiency of email data extraction, enabling businesses to automate tasks and reduce costs. This approach can be particularly beneficial for ecommerce companies that process large volumes of email data.

✅ Practical Steps

1. Prepare training data in the Amazon Bedrock conversation format with email content as input and extracted entities as output.
2. Upload training data to Amazon Simple Storage Service (Amazon S3).
3. Create a custom model fine-tuning job using Amazon SageMaker AI and Amazon Nova recipes.

Want the full story? Read the original article.