Anatomy of a training job
Lesson 1 of 4 in Training and Tuning on SageMaker AI.
A SageMaker training job is a contract, not a server. You declare what to run (an algorithm container — built-in, framework + your script, or your own image), what to feed it (input channels pointing at S3), what to run it on (instance type and count), and where results go (an S3 output path). SageMaker AI provisions the fleet, mounts the data, runs the container, streams telemetry to CloudWatch, uploads the model artifact, and tears everything down. You pay for the seconds the fleet lived.
| Path | What lands there |
|---|---|
| Your S3 channels, one directory per channel (train, validation, …) |
| The hyperparameters you passed, as JSON strings |
| Everything written here is uploaded to S3 as model.tar.gz when the job ends |
| Default local checkpoint path, synced to your S3 checkpoint URI during training |
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.