Amazon has announced new AI data preparation capabilities for SageMaker AI, aimed at the slowest part of custom model work: getting training and evaluation data into shape. They are generally available in US East (N. Virginia) and US West (Oregon).
What was added
The release is a set of standalone building blocks that can be used singly or chained:
- Auto-labeling. Labels a dataset starting from unlabeled rows, a small labeled seed set, or nothing. The system comes with a range of prebuilt task types, such as sorting items into categories, pulling out named entities, answering questions by locating spans within a text, assigning ordered ratings, and condensing content, and people may also create tasks tailored to their own needs. A short labeling spec is validated against your own examples before the full run. For tasks with fixed answer sets, several models label each row, and how much they agree becomes a confidence score.
- Human review. Rows under your confidence threshold go to a queue in SageMaker Studio, built on the open-source Label Studio Community Edition. External subject experts can reach it through a link, and their corrections feed the next labeling pass.
- Synthetic data generation. Produces training, evaluation and test data from a specification, grounded in your seed data and documents. The service writes a code-based pipeline you can read, change and reuse.
- Data quality evaluation. Tests a dataset against your business rules, built-in metrics and LLM-as-a-judge metrics, and flags duplicates, near-duplicates, label noise and answer leakage.
- Data transformation. Cleans and reshapes data for the target training method, such as SFT, DPO or reinforcement learning, via plain-language instructions, built-in operations like dedupe, merge, split and filter, or your own Python.
How you use it
You can call the SageMaker Python SDK directly, or give a coding agent the AI-ML skill and let it pick and run capabilities in SageMaker Studio JupyterLab, Kiro or a local IDE. The SDK takes a specification as input, and AWS recommends drafting that specification with the agent first.
AWS describes one example for support teams: supply a product guide, ask for questions and reference answers, and require that answers about a factory reset keep the warning that saved settings are erased. The team generates a small sample, checks it against the guide, runs a quality assessment, has an expert review ambiguous rows, and repeats until the set is ready to test an assistant.
Control and data location
The service runs under an IAM execution role you control. Source data, generated datasets and labels remain in your own S3 buckets in the region you pick.
Splunk is the one customer named. A director of engineering for its Agent Observability product says synthetic generation helps it scale evaluation models and produces representative datasets quickly. That is a vendor-supplied testimonial, and the source gives no figures on speed or accuracy.
What the source does not say
The announcement gives no pricing and no timeline for more regions, so cost and wider availability are unconfirmed.