Enterprise
AWS launches AI data prep tools for SageMaker model workflows
New capabilities handle synthetic data generation, auto-labeling, quality checks, and human review without custom pipeline code.
Key takeaways
- AWS released five AI data preparation capabilities in SageMaker: auto-labeling, human review, synthetic data generation, quality evaluation, and transformation.
- Auto-labeling uses multiple models to label each record and generates confidence scores based on their agreement level.
- Teams can chain capabilities together—for example, generating synthetic data, transforming it for specific training formats, validating quality, then training models.
- Splunk used the synthetic data generation to create evaluation datasets for task-specific models in its Agent Observability product.
- All data preparation operations run in customer-controlled AWS regions with data staying in customer-owned S3 buckets.
Amazon Web Services has released a suite of AI data preparation capabilities within Amazon SageMaker designed to address one of machine learning's most persistent bottlenecks: assembling clean, task-specific training and evaluation datasets.
The new tools, now generally available in US East (N. Virginia) and US West (Oregon) regions, provide five core functions through Python SDK APIs and an AI-ML agent skill: auto-labeling, human review workflows, synthetic data generation, data quality evaluation, and data transformation.
Why it matters
Data preparation typically consumes the majority of time in ML projects, yet most teams still cobble together labeling tools, generation scripts, and validation jobs manually. For organizations building specialized models—especially in regulated industries where data is scarce or privacy-constrained—the ability to generate synthetic examples, validate quality systematically, and incorporate expert review without custom engineering can compress weeks of work into days. Smaller models trained on high-quality, task-specific data frequently match or outperform larger general-purpose models at lower inference costs.
How the capabilities work together
The auto-labeling function supports classification, named entity recognition, extractive question answering, ordinal rating, and summarization tasks. Teams describe the task in a labeling specification, which the system validates against sample data before processing the full dataset. For tasks with discrete answer sets, multiple models label each record and their agreement level produces a confidence score.
Records falling below a confidence threshold route to a review queue in SageMaker Studio, powered by the open-source Label Studio Community Edition. External subject matter experts can access the review interface through partner app links. Reviewer corrections feed back into the auto-labeling process, creating an iterative improvement loop.
The synthetic data generation capability creates training, evaluation, and test datasets from specifications grounded in seed data and documents. The service produces a code-based generation pipeline that teams can inspect, modify, and scale.
Data quality evaluation checks datasets against custom business rules, built-in metrics, and LLM-as-a-judge assessments. It identifies duplicates, near-duplicates, label noise, and answer leakage—problems that often surface only after expensive training runs fail to meet expectations.
Data transformation handles cleaning and reshaping for specific training techniques. Teams can describe transformations in natural language to generate code, use built-in operations like dedupe and filter, or run custom Python at scale.
Real-world application
Splunk, a Cisco company, used the synthetic data generation capability for its Agent Observability product. "Our ability to train robust GenAI Eval models relies on access to high-quality, diverse datasets which are hard to curate for very specific tasks," said Yash Sheth, Director of Engineering for Splunk Agent Observability. The company reported generating statistically representative synthetic datasets reliably while preserving underlying patterns required for model performance.
Implementation options
Teams can work through the AI-ML skill with coding agents in SageMaker Studio JupyterLab, Kiro, or local IDEs, or invoke Python SDK APIs directly from existing code. All operations run under IAM execution roles controlled by the customer, with source data, generated datasets, and labels remaining in customer-owned S3 buckets within their chosen AWS region.
These details were first reported by AWS in a Machine Learning Blog post announcing the capabilities.