SageMaker can now label, generate and audit training data from your notebook
Amazon SageMaker AI adds auto-labeling with a model jury, synthetic data generation, quality checks and format conversion, usable from the Python SDK or a coding agent.

Amazon has added AI data preparation capabilities to SageMaker AI, covering auto-labeling, synthetic data generation, quality checks and format conversion. According to the AWS announcement, they are generally available in two US Regions.
Five building blocks, usable alone or chained
AWS describes each capability as a standalone piece that can be used on its own or combined. Auto-labeling can start from raw unlabeled data, from a handful of labeled examples, or from nothing but a task description. Built-in task types are classification, named entity recognition, extractive question answering, ordinal rating and summarization, and custom tasks can be defined. The user writes a short labeling spec, and AWS says it is tested on the user’s own examples before the whole dataset gets labeled.

Where the possible answers are a closed set, as in classification, entity recognition and extractive question answering, a jury of several models labels each row. Their level of agreement becomes the row’s confidence score. Rows below a chosen threshold go to a review queue in SageMaker Studio, built on the open-source Label Studio Community Edition. Outside subject matter experts can reach it through a link in the partner app box, and their corrections feed back into the next auto-labeling pass.
Synthetic data generation produces training, evaluation and test data from a specification, optionally grounded in the customer’s own seed data and documents. The service generates a code-based pipeline that can be reviewed, modified and reused. Data quality evaluation checks datasets against business rules, built-in metrics and LLM-as-a-judge metrics, finds duplicates and near-duplicates, and flags problems such as label noise or answer leakage. Data transformation reshapes data for a target technique through plain-language descriptions that generate code, built-in operations such as dedupe, merge, split and filter, or custom Python.
Who can use it, and where the gaps are
The tools are available in US East (N. Virginia) and US West (Oregon). There are two routes in: the SageMaker Python SDK, or the AWS AI-ML skill, which lets a coding agent call the APIs from SageMaker Studio JupyterLab, Kiro or a local IDE. The SDK takes a specification as input, and AWS recommends creating that specification with the agent first.
Runs use an IAM execution role the customer controls. Source data, generated datasets and labels stay in the customer’s own S3 buckets in the chosen Region, AWS says. A companion What’s New post adds that teams can start from a few examples, production traces, or data in S3, Athena or the AWS Glue Data Catalog, and that the scope is text-based model customization and traditional ML.
Neither page states pricing, and the blog points readers to the SageMaker AI documentation for details. Nor does AWS publish accuracy figures for the jury approach or for the synthetic data, so the claims about speed and quality are the company’s own.
“We’ve been able to generate statistically representative synthetic datasets reliably and fast, while preserving the underlying patterns our models depend on.”
That is how Yash Sheth, Director of Engineering for Splunk Agent Observability, described the synthetic data tool in AWS’s post. It is the only customer evidence offered, and it comes without measurements.
A pipeline from raw documents to training
AWS sketches two chains. For a golden evaluation set, raw data is auto-labeled from a small seed set, confident rows are accepted, and humans fix the rest. For model training, the steps run from synthetic generation to transformation into the format the technique expects (SFT, DPO or RL), then quality validation, then SageMaker AI model customization. The aim, in AWS’s words, is to catch duplicates and leakage before compute is spent on training.
AWS’s worked example is an evaluation set for customer support assistants, where answers about a factory reset must keep a warning that saved settings are erased. The loop is to define requirements, generate a small sample, assess it with expert review, then revise and repeat.
What to watch as data prep becomes agent-driven
The design choice worth tracking is that the specification, not the dataset, is the main artifact. Specs and generated pipelines are inspectable and editable, which fits how AWS is pushing coding agents into its tooling. Our earlier report on AWS sign-up for Claude Code and Cursor showed the same direction: agents as the first-class users of AWS services.
The data matters because technique choices increasingly depend on it. Our coverage of Nvidia’s Nemotron recipe reported reinforcement learning lifting drug-metabolism accuracy 17% past plain fine-tuning, a result that needs task-specific data in the right format. AWS names SFT, DPO and RL as targets for its transformation step.
Several things would confirm or weaken AWS’s pitch. Published pricing will show whether jury labeling, which runs several models per row, is cheaper than human annotation at scale. Expansion beyond the two US Regions would matter to teams with data residency rules. Independent comparisons of jury confidence scores against human labels would test whether agreement really tracks correctness. Until then, the release is a convenient packaging of known techniques inside SageMaker, with its value resting on AWS’s own description.
Sources
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.



