The Pulse
AWS adds AI data preparation tools to SageMaker AI
SageMaker AI users can generate synthetic data, label records and check datasets for model training and evaluation. The capabilities are generally available in two AWS regions and can be used through the Python SDK or a coding agent.

SageMaker AI users can now generate synthetic examples, label records and assess datasets for model training and evaluation through new data preparation tools. Amazon Web Services said the capabilities can be used through Python APIs or a coding agent, with specifications and code that users can inspect and change.
The capabilities are generally available in the US East (N. Virginia) and US West (Oregon) AWS regions. They can be used separately or combined in a workflow, from preparing raw data through to training a model.
Labels and human review
The auto-labeling tool works from an unlabeled dataset, a small labeled seed set or no labels. Built-in tasks include classification, named entity recognition, extractive question answering, ordinal rating and summarization; users can also define custom tasks. A labeling specification is checked against the user’s examples before the full dataset is labeled.
For tasks with a fixed set of possible answers, including classification, multiple models label each row and their agreement determines a confidence score. Records below a user-selected threshold go to a review queue in Amazon SageMaker Studio.
The review app uses the open-source Label Studio Community Edition, and outside subject matter experts can access it through a link in Studio. Reviewer corrections feed back into auto-labeling, while users can generate consensus metrics and labeling reports.
Generation and quality checks
Synthetic data generation produces training, evaluation and test data from a specification. Users can ground the output in their own seed data and documents, then inspect, modify and reuse the generated pipeline code.
Quality evaluation checks data against business rules, built-in metrics and metrics that use a large language model as a judge. It can find duplicates and near-duplicates and surface issues such as label noise or answer leakage.
A separate transformation capability can clean and reshape data using generated code, built-in operations such as merging, splitting and filtering, or the user’s own Python code. AWS describes a workflow in which generated examples are transformed for a training technique, checked for quality issues and then used with SageMaker AI model customization.
Running the jobs
Users can call individual capabilities through the SageMaker Python SDK or work through a coding assistant using the AWS AI-ML skill. The assistant can help draft specifications and coordinate steps in SageMaker Studio JupyterLab, Kiro or a local development environment.
Jobs that produce data register it as a dataset in the user’s account, ready for the next step or for model training. The assistant asks for approval before running a job or creating resources, according to AWS documentation.
The tools require version 3.30.0 or later of the SageMaker Python SDK, which requires Python 3.10 or later. Jobs also need an execution role and a default Amazon S3 bucket in a supported region. Source data, generated datasets and labels stay in the user’s S3 buckets in the AWS region they choose.
Sources
Companies in this story
Related stories
- AWS adds model leaderboard to SageMaker JumpStart
- AWS details OpenFold3 research preview deployment on SageMaker
- AWS reports full fine-tuning results for SageMaker serverless
- AWS adds stalled-job recovery and checks to SageMaker HyperPod
- AWS details per-user cost tracking and limits for Bedrock AI tools