AWS enables fine-tuning of Amazon Nova models for email extraction, achieving 94.77% accuracy with 50% cost reduction

AWS published detailed guidance on fine-tuning Amazon Nova models for automated email data extraction, demonstrating significant improvements in accuracy and cost efficiency. The approach uses Amazon SageMaker AI with Parameter-Efficient Fine-Tuning (PEFT) through Low-Rank Adaptation (LoRA).

Production results from Parcel Perform

Parcel Perform, an AI delivery experience platform processing millions of emails daily, collaborated with AWS Generative AI Innovation Center to optimize Nova models for extracting structured data from diverse email formats. According to AWS, the fine-tuned Nova Micro model achieved 94.77% extraction accuracy on the testing dataset—an improvement of 16.6 percentage points over baseline performance.

The fine-tuned Nova Micro reduced inference latency by more than 30% and cut costs by 50% compared to Parcel Perform's previous model. The company has moved the solution into production for e-commerce logistics operations.

Technical implementation

The solution uses supervised fine-tuning with LoRA on Amazon SageMaker AI Training, then deploys the customized model to Amazon Bedrock with on-demand inference priced per token. AWS tested two training datasets: 1,300 samples and 4,900 samples to evaluate how data volume impacts performance.

Key training parameters include:

• Base models: amazon.nova-lite-v1:0:300k and amazon.nova-micro-v1:0:128k
• Context length: 8,192 tokens on g5/g6 instances, 32,768 tokens on p5 instances
• Global batch size: 64
• Training epochs: 2
• LoRA alpha: 32
• LoRAPlus learning rate ratio: 8.0

Training data must follow Amazon Bedrock's conversation schema format with email content as user input and extracted entities as assistant responses in JSON format.

Deployment options

With PEFT, models deploy to Amazon Bedrock using on-demand inference. Full-rank supervised fine-tuning supports deployment through either Provisioned Throughput on Amazon Bedrock or a SageMaker AI endpoint. The approach addresses common challenges including model hallucinations, confusion between similar data types (order numbers versus tracking numbers), and high token costs when processing HTML-formatted emails.

Requirements

Developers need an AWS account with permissions for Amazon Bedrock and SageMaker AI, an IAM service role for model customization, an S3 bucket for training data storage, training data in JSONL format, and sufficient Service Quotas for chosen instance types. AWS provides Amazon Nova recipes—YAML configuration files specifying base model names, training hyperparameters, and optimization settings.

What this means

This marks AWS's first detailed public guidance on fine-tuning its Nova model family for production use cases. The 94.77% accuracy with simultaneous 50% cost reduction and 30% latency improvement demonstrates that smaller, fine-tuned models can outperform larger base models on domain-specific tasks. The LoRA approach requires minimal training data (1,300-4,900 samples) compared to full fine-tuning, making specialized model development more accessible. For organizations processing high volumes of structured data extraction, this approach provides a clear path to improving accuracy while controlling inference costs.