Quick answer
Amazon SageMaker AI supports serverless full fine-tuning for selected open-weight models. Start from the JumpStart and Models page in SageMaker Studio or use the SageMaker Python SDK. AWS manages training infrastructure provisioning and orchestration, but you still own model selection, IAM, data quality, evaluation, cost limits, and deployment approval.
Use this sequence:
- choose an exact supported model, technique, and Region;
- configure a SageMaker domain and execution role;
- create versioned training and evaluation assets;
- submit the customization job in Studio or through the SDK;
- monitor metrics, logs, artifacts, and failures;
- evaluate an immutable holdout;
- register and deploy only an approved model version.
Prerequisites
AWS documents SageMaker Studio domain access, a current AWS CLI, configured credentials, and model-customization permissions. The managed AmazonSageMakerModelCustomizationCoreAccess policy is AWS’s recommended basic option; security teams should still review whether a narrower custom role fits the workload.
Confirm access to the intended S3 buckets, model packages, MLflow resources, training jobs, evaluation jobs, and selected deployment path. Keep training data, outputs, and logs in the approved account and Region boundary.
Configure the job
| Field | Decision |
|---|---|
| Base model | Pin the AWS model ID and source license |
| Training type | Full fine-tuning or LoRA |
| Technique | SFT, DPO, RLVR, or RLAIF only where the model matrix supports it |
| Data | Versioned S3 training and evaluation assets |
| Output | Approved S3 path and model package group |
| Hyperparameters | Reviewed defaults or measured changes |
| Evaluation | Holdout, metrics, scorer, human review, and failure criteria |
The Studio flow begins from a model card and Customize model. The SDK exposes technique-specific trainers. Do not copy a DPO or LoRA example into a full fine-tuning job without checking the current SDK class and TrainingType documentation.
Monitor and approve
After submission, review the job page for training configuration, live metrics, logs, and generated artifacts. SageMaker creates logged-model records for successful jobs and can retain checkpoint or metadata records for failures. Treat a completed job as an artifact, not an approval.
Evaluate the custom model on an untouched dataset, compare it with the base model, inspect regressions and unsafe outputs, and record lineage before deployment. Use the data and evaluation guide and cost and operations guide before scaling.
Frequently asked questions
Does SageMaker serverless full fine-tuning require provisioning GPU instances?
AWS says SageMaker model customization automatically provisions and cleans up training infrastructure, so you submit the job without selecting and managing training instances yourself.
Can I submit a SageMaker model customization job from code?
Yes. AWS documents both the SageMaker Studio workflow and programmatic submission through the SageMaker Python SDK.
What should I verify before starting a full fine-tuning job?
Verify the exact model and technique support, Region, Studio domain, execution role, dataset and output locations, evaluation plan, budget, model terms, and deployment target.
Official sources
- AWS announcement: SageMaker serverless model customization supports full fine-tuning
- SageMaker model customization guide
- SageMaker open-weight job submission
- SageMaker model customization prerequisites
Source check: August 4, 2026. Verify current SDK APIs, permissions, model and technique support, Regions, quotas, prices, and terms before submitting a job.