AWS Details AI Product Tagging System Built With SageMaker Serverless Model Customization

AWS published a technical walkthrough describing how to build an AI-powered product tagging system using Amazon SageMaker serverless model customization, according to a blog post from AWS authors Linpo Guo, Ray Wang, Josh Chiu, and Kanwaljit Khurmi.

The problem AWS describes

AWS says retail catalogs rarely arrive with clean, structured attributes. Product names, descriptions, and category paths come from many sources and change continuously, the company says, and while search, recommendations, and catalog navigation depend on consistent tags, manually applying those tags across thousands of stock keeping units (SKUs) is slow and hard to keep consistent.

AWS says a general-purpose frontier model can generate tags through prompt engineering, but argues that when a tagging workflow has a narrower, stable objective — returning the right attributes in a fixed schema — customizing a smaller open-weight model can be a better fit. According to AWS, this approach lets a team teach the model the schema directly, optimize the trade-off between missing tags and unnecessary tags, and avoid paying for broad model capabilities the workflow doesn't need on every request.

What AWS built

In the walkthrough, AWS says it customizes the Qwen3-8B open-weight model in two stages: first with supervised fine-tuning (SFT), then with reinforcement learning with verifiable rewards (RLVR) using an approach called Group Relative Policy Optimization (GRPO).

AWS states that Amazon SageMaker serverless model customization manages the training capacity for this process. When no compute configuration is supplied, the company says, Amazon SageMaker selects and releases the training capacity for the customization job itself. The optimized model is then deployed separately, AWS says, to Amazon SageMaker Asynchronous Inference for batch-oriented catalog enrichment.

AWS distinguishes this approach from an earlier Qwen3-8B example in its amazon-sagemaker-examples repository, which it says used Amazon SageMaker Training Jobs (SMTJ) with customer-selected GPU instances and custom training images. This new walkthrough instead uses the Amazon SageMaker Python SDK v3 serverless customization trainers, which AWS names as SFTTrainer and RLVRTrainer.

Prerequisites AWS lists

AWS lists several resources and permissions needed before following the walkthrough:

• Amazon SageMaker AI permissions to manage serverless customization jobs, AI Registry datasets and evaluators, model package groups, models, endpoints, and asynchronous inference, plus iam:PassRole where required
• Amazon S3 access to read and write the source catalog, transformed training data, model artifacts, and asynchronous inference requests and outputs
• Amazon ECR access, which AWS says is required only if a custom vLLM inference image is built and hosted, noting that serverless SFT and RLVR do not require a custom training image
• Confirmation that the chosen AWS Region and model/technique combination supports Qwen3-8B SFT and RLVR
• Sufficient hosting quota for ml.g6.2xlarge instances and the endpoint count used for the asynchronous endpoint

For data, AWS says the walkthrough uses the public Amazon Sales Dataset, or alternatively a private catalog transformed into the same prompt/target structure.

Source: AWS Machine Learning Blog, "Build an AI-powered product tagging system with Amazon SageMaker serverless model customization," published September 15, 2026.