Amazon Nova models fine-tuned with Amazon SageMaker AI have demonstrated significant improvements in email data extraction for Parcel Perform, an e-commerce logistics company. The fine-tuned models achieved up to 94.77% extraction accuracy on a testing dataset, an improvement of up to 16.6 percentage points over the baseline. This enhancement helped reduce inference latency by more than 30 percent and cut costs in half compared to Parcel Perform’s previous model. The solution also matched or exceeded the performance of the fine-tuned Nova Lite model at a lower cost. These results enabled Parcel Perform to deploy the solution in production, improving its e-commerce logistics operations.
The fine-tuning process involved supervised learning with Parameter-Efficient Fine-Tuning (PEFT) through Low-Rank Adaptation (LoRA). This approach allows models to be customized effectively with limited training data while maintaining computational efficiency. Amazon SageMaker AI was used to create the fine-tuning job, which ran on GPU-powered instances and automatically stopped after training completed. Training data was uploaded to Amazon S3, and the model was deployed using Amazon Bedrock with on-demand inference. The custom model fine-tuning used Amazon Nova recipes, which are YAML configuration files that specify how to run the model customization job, including base model names, training hyperparameters, and optimization settings.
Parcel Perform collaborated with the AWS Generative AI Innovation Center (GenAIIC) to optimize Nova models for its specific needs. The team scoped a project to improve multiple metrics: accuracy, latency, and cost. Le Vy, AI Team Lead at Parcel Perform, reported the results of the fine-tuning process, which included the performance improvements and cost reductions achieved through the customization. The collaboration allowed for concurrent improvements across these metrics, demonstrating the effectiveness of the fine-tuning approach.
Source: awsml
