aws/amazon-sagemaker-examples

What it solves

This repository provides a comprehensive collection of Jupyter notebooks that demonstrate how to use Amazon SageMaker to build, train, deploy, and operationalize machine learning models. It also introduces SageMaker-Core, a new Python SDK designed to simplify the management of SageMaker resources like training jobs, models, and endpoints through an object-oriented interface.

How it works

The project is organized by ML capability, following the standard machine learning lifecycle:

• Training: Examples of managed training at scale, including distributed training and managed spot training.
• Model Customization: Demonstrations of fine-tuning techniques such as SFT, DPO, RLVR, RLAIF, and continued pre-training.
• Model Evaluation: Tools for scoring models using benchmarks, custom scoring functions, or LLMs as judges.
• Inference: Guidance on deploying models via real-time, serverless, and asynchronous endpoints, as well as batch transform.
• MLOps: Examples of implementing CI/CD, lineage tracking, and experiment tracking using SageMaker Pipelines and MLflow.

Who it’s for

ML practitioners and developers who want to implement machine learning workflows on AWS, specifically those seeking full customization of AWS primitives for their ML workloads.

Highlights

• SageMaker-Core SDK: An object-oriented Python SDK that replaces manual parameter specification with resource chaining.
• Full Lifecycle Coverage: Covers everything from training and been fine-tuned to deployment and MLOps.
• Linguistic Diversity: Supports a wide range of fine-tuning techniques (SFT, DPO, RLVR).
• Integrated Environment: Notebooks are automatically loaded into SageMaker Notebook Instances for immediate use.