Demo Fine tuning a Pre trained Model with SageMaker JumpStart
Guide to fine-tuning pre-trained models using AWS SageMaker JumpStart, covering artifacts, compute, hyperparameters, training workflow, and LoRA customization
Fine-tuning adapts a pre-trained model to your specific dataset and task so you avoid training from scratch. It is typically faster, less expensive, and often yields better results when labeled data is limited.
Why start from pre-trained checkpoints? Organizations and open-source projects have already trained large models and released checkpoints with solid baseline performance. Using JumpStart models lets you leverage that prior compute and expertise—then adapt the checkpoint to your domain, labels, or task through fine-tuning.
Pre-trained checkpoints save time and GPU compute. Fine-tuning adapts an existing model to new labels, domains, or tasks without starting from random initialization.
Some JumpStart entries are inference-only (evaluation/deployment artifacts), while others include training workflows. For example, a Table Transformer Detection model in JumpStart may be offered primarily for evaluation and deployment:
When a model supports training, SageMaker JumpStart provides a guided job-creation workflow. You will supply three primary artifacts and a compute configuration.
Required artifacts and where to put them:
Artifact
Description
Example
Input model artifact
The pre-trained checkpoint (weights, config, tokenizer). Typically an S3 prefix.
s3://my-bucket/models/pretrained/
Training dataset
Your labeled data for fine-tuning. Format depends on model (CSV, JSONL, TFRecord, images, etc.).
s3://my-bucket/data/training/
Output artifact location
Where SageMaker writes fine-tuned checkpoints and metadata.
s3://my-bucket/models/finetuned/
Think of artifacts as serialized model state. Each training run produces new files (updated weights, optimizer state, tokenizer changes, training logs and metrics).You also select compute for training. JumpStart exposes recommended defaults such as ml.m5.xlarge (CPU) or GPU instances like ml.p3.2xlarge. Choose instance type and count based on model size, memory/GPU needs, and budget.
Hyperparameters — what to configure and why
For classical models (e.g., linear classifiers) you’ll often choose penalty/regularization, tolerance, and solver parameters.
For deep models, tune learning rate, batch size, number of epochs, weight decay, and optimizer type.
Example hyperparameter block for a linear classifier (illustrative):
{ "penalty": "l2", "C": 1.0, // inverse of regularization strength (scikit-learn style) "tol": 1e-4, "l1_ratio": 0.15 // used only if penalty == 'elasticnet'}
Notes:
Some libraries (e.g., scikit-learn) use C as the inverse of regularization strength; other frameworks use alpha or lambda.
Always start with default values and run small experiments to find good hyperparameters before scaling up.
After launching the job, monitor training status, logs, and metrics in Studio’s training jobs view.
JumpStart lists created jobs (including JumpStart-created training jobs and other job types) in the Studio Jobs view. From there you can inspect job state, stream logs, and follow links to input/output artifacts and metrics.For large generative models, JumpStart also supports model customization strategies. Low-Rank Adaptation (LoRA) is a widely-used technique that injects low-rank update matrices into transformer layers instead of modifying full model weights. LoRA drastically reduces GPU memory usage and the storage required for fine-tuned weights by saving only the low-rank adapters.
Fine-tuning large models (even with LoRA) can be expensive in GPU time and data transfer. Monitor costs, run small experiments first, and cap runtimes or instance counts for early trials.
Quick checklist before you submit a JumpStart fine-tuning job:
Step
Action
1
Confirm the chosen JumpStart model supports training/fine-tuning.
2
Upload or reference the pre-trained checkpoint in S3 (input model artifact).
3
Upload or reference the training dataset in S3.
4
Provide an output S3 location for resulting artifacts.
5
Choose instance types and count appropriate for the model.
6
Set hyperparameters (learning rate, batch size, regularization or C, tol, l1_ratio when applicable).
7
Submit the training job and monitor logs, metrics, and outputs.
This article explained how to prepare and submit a fine-tuning job in SageMaker JumpStart, how artifacts map to S3 locations, the compute and hyperparameter choices you’ll configure in Studio, and model customization options such as LoRA.
