Fine-tune Amazon Nova models for accurate email data extraction

This article demonstrates how fine-tuning Amazon Nova models using Amazon SageMaker AI enables accurate entity extraction from ecommerce emails while reducing costs and hallucinations.

• Fine-tuned Nova Micro achieved 94.77% extraction accuracy, improving 16.6 percentage points over baseline
• Parameter-Efficient Fine-Tuning (PEFT) with LoRA reduces inference latency by over 30% and cuts costs by 50%
• Smaller Nova Micro model outperformed larger Nova Lite after fine-tuning on domain-specific tasks
• Meaningful accuracy gains achieved with just 1,300 training samples across 25 entities
• Deploy fine-tuned models to Amazon Bedrock with on-demand, token-based pricing
• Solution addresses hallucinations and confusion between similar data types like order and tracking numbers

Fine-tuning Amazon Nova models provides a cost-effective, production-ready approach for specialized entity extraction tasks without requiring massive datasets or dedicated infrastructure.