News

AWS documents SageMaker AI data preparation workflows

AWS's SageMaker Python SDK V3 documents coordinated workflows for synthetic-data generation, quality checks, transformations, automatic labeling and review routing.

D
Nov 17, 2026 · 2 min read

AWS’s SageMaker Python SDK V3 documentation describes data-preparation capabilities for synthetic-data generation, dataset evaluation, transformation and automatic labeling. Developers can use the interfaces to coordinate the assembly and inspection of training or evaluation datasets. The documented functions still run as separate jobs, however, rather than as one end-to-end server-side pipeline.

The documentation describes DataPreparer as a shared entry point for generation, evaluation and transformation. Developers call generate, evaluate or transform with related configuration patterns. In the documented synthetic-data workflow, they create a sample, refine the specification and then run full generation. The evaluation workflow imports a dataset, scores it and produces a scorecard with per-record annotations.

According to the API reference, generation can start from a task specification and an optional seed dataset, while evaluation writes quality annotations for individual records. Transformation operators include deduplication, dataset splitting and merging, schema conversion, quality filtering and custom code. The service currently accepts exactly one transformation operation per job, so a multi-step workflow requires developers to chain jobs and pass one output dataset ARN into the next.

Automatic-labeling configurations can pair gold evaluation data with a scoring target and route low-confidence rows into a named review dataset. The resulting summary can show counts for labeled, accepted, routed, failed, missing, human-labeled and spot-checked rows. It can also expose held-out accuracy, a 95% lower bound, acceptance thresholds, warnings and output dataset identifiers. For classification, a review-only option checks seed labels against the specification without labeling new rows or registering a dataset.

These outputs make the workflow inspectable, but they do not show that the resulting data is accurate or unbiased. AWS’s responsible-AI guidance recommends comparing generated labels with human judgment, checking synthetic data for bias and edge-case coverage, comparing relevant statistical properties with real data, and documenting limitations and failure modes. The inspected sources contain no independent benchmark, customer result, adoption figure, time saving, cost reduction or accuracy measurement for the documented capabilities.

Labeling and human review are not new to SageMaker. Ground Truth already used active learning and human workers for supported image and text tasks. AWS says Ground Truth is closed to new customers and that it does not plan to introduce new features. Its documented accuracy thresholds are expectations based on human-annotated validation sets, not independent measurements of the newly documented data-preparation tools.

AWS prices synthetic-data generation by input tokens for grounding samples and contextual documents, plus output tokens for generated datasets, according to the SageMaker pricing page. AWS Labs lists a SageMaker AI plugin for coding assistants, but the inspected repository does not establish that it covers every newly documented data-preparation API.

More news