Urgent.News

What's breaking now, across thousands of outlets.

AI

Announcing AI data preparation capabilities in Amazon SageMaker: Build high-quality datasets for model development workflows

Amazon SageMaker AI now offers AI data preparation capabilities for building high-quality training and evaluation data. Generate synthetic examples, assess and improve datasets, auto-label at scale, and bring in expert review through Python SDK APIs and an agent skill, all from your own notebook or IDE.

Recent enhancements to Amazon SageMaker AI now provide teams building generative AI solutions with tools to prepare high-quality datasets for model development. Crafting trained and evaluated data is a crucial first step, as model accuracy hinges on precise, well-labeled, and representative information. While smaller, task-specific models can sometimes rival larger general-purpose models in performance, the data preparation phase remains a significant bottleneck.

Poor data quality, such as duplicates, label noise, and answer leakage, often goes unnoticed until a training run produces disappointing results. Consequently, data scientists spend extensive time cleaning and refining data, delaying model development.

To address this challenge, Amazon SageMaker AI now offers capabilities for generating synthetic examples, assessing and improving existing datasets, and incorporating expert feedback into the data preparation process. These features are accessible via Python SDK APIs and the AWS AI-ML agent skill, allowing users to work within their preferred notebooks or integrated development environments (IDEs).

With these standalone building blocks, users can chain them together to create customized workflows tailored to their specific needs.

Auto-labeling enables labeling data at scale from unlabeled datasets, small labeled seed sets, or no labels at all. Predefined task types cover classification, named entity recognition, extractive question answering, ordinal rating, and summarization, with the ability to define custom tasks. Each row's confidence score is determined by a panel of several models in the case of fixed-answer tasks, while human review of low-confidence rows is facilitated through Amazon SageMaker Studio and the Label Studio Community Edition.

Corrections from these reviews are fed back into the auto-labeling process, generating consensus metrics and labeling reports.

Synthetic data generation allows the creation of training, evaluation, and test data from a specification, grounding the output in the user's seed data and documents. The resulting code-based data generation pipeline can be reviewed, modified, reused, and scaled as needed. Data quality evaluation checks datasets against business rules, built-in metrics, and LLM-as-a-judge metrics, identifying duplicates and near-duplicates, as well as less obvious issues like label noise and answer leakage.

Data transformation cleans and reshapes data for the target training technique, with users describing the changes in plain language, allowing the system to generate the corresponding code. Operations such as deduplication, merging, splitting, and filtering are available, along with the option to run custom Python code at scale. Users can work directly with the SageMaker Python SDK or utilize the AI-ML skill in their agent toolkit, which invokes the necessary APIs in accordance with data science best practices.

By combining these AI data preparation capabilities, teams can build golden evaluation datasets by auto-labeling raw data from a small seed set. High-confidence rows are automatically accepted, while low-confidence rows are reviewed by human experts, whose corrections improve subsequent labeling passes. This results in a trusted, human-validated dataset for benchmarking models against.

Furthermore, the process can generate task-specific examples from raw documents and seed data, reshape the data into the required format for training techniques like supervised fine-tuning, direct preference optimization, or reinforcement learning, and validate the dataset for duplicates, leakage, and rule violations before training begins.

Ultimately, these enhancements streamline the data preparation process, enabling data scientists to spend more time building models and driving real enterprise value.

Written by urgent.news from AWS Machine Learning's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

This story

This is one outlet's version. Read the fullest account.

Every outlet that reported this, and when →

Read the original at aws.amazon.com →

More in AI

KERV.ai, Macmillan Publishers and Chomps are among this year’s Work/Life Awards finalists

This year’s Work/Life Awards finalists reflect a landscape where companies are rethinking workplace culture as a driver of business performance, employee engagement and innovation.

  • KERV.ai recognized in Most Innovative Culture for employee input and collaboration
  • Macmillan Publishers nominated for Most Committed to Social Good
  • Employee volunteering at Macmillan increased by 60% in 2025

More from Tuesday 17 November →