AWS details full fine-tuning in SageMaker serverless customization
AWS explains how SageMaker AI customers can update every model weight through its managed customization service, contrasting the workflow with LoRA and publishing company-run benchmark results.
AWS has published detailed guidance on running full fine-tuning through SageMaker AI Serverless Model Customization. The managed workflow lets customers update all of a model’s weights without provisioning GPUs, configuring clusters or setting up distributed training. The capability did not launch in November: AWS announced its availability on August 3.
The service offers full fine-tuning alongside Low-Rank Adaptation, or LoRA, but the two methods work differently. AWS’s model customization documentation describes full fine-tuning as updating every model weight, which requires more compute but allows deeper customization. LoRA instead trains a smaller set of adapter weights while leaving the pretrained weights frozen, a design that lowers training cost and time. The original LoRA paper describes those adapters as trainable low-rank decomposition matrices inserted into the model’s layers.
For its November 18 technical post, AWS compared the two approaches using supervised fine-tuning on a Qwen3-4B base model and the Magicoder-Evol-Instruct-110K coding dataset. The company split the data into 99,000 training records, 5,500 validation records and 5,500 held-out test records, and said both methods received the same prompts, targets and evaluation code.
AWS reported 90.0% semantic correctness in an LLM-as-Judge evaluation at what its post calls the “100K+” training scale; elsewhere, the same post describes a 99,000-record training split from the 110,000-record dataset. That figure comes from AWS’s experiment and has not been independently replicated in the evidence reviewed for this story. In a separate comparison, the company reported that LoRA scored 81.5% on the same type of evaluation, ahead of full fine-tuning at 79.5%. AWS said full fine-tuning is not universally better.
The service’s built-in evaluation workflow supported three methods used in AWS’s test: an LLM-as-Judge assessment using Amazon Nova Pro, a PRIMECODE custom scorer based on text-similarity measures, and MMLU for tracking general-knowledge retention. AWS said the resulting models were automatically registered in SageMaker Model Registry as versioned Model Packages for review or deployment.
AWS warned that changing every parameter makes full fine-tuning sensitive to hyperparameter choices. In the company’s experiments, learning rates of at least 5e-5 pushed MATH scores down to 2–3% and MMLU accuracy to 3.6%, results AWS presented as evidence of catastrophic forgetting, where task-specific training overwrites broader capabilities.
At launch, AWS said serverless full fine-tuning supported more than 25 open-source models and was available in US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo) and Europe (Ireland). Customers could start customization jobs in SageMaker Studio or through the SageMaker Python SDK.
More news

Shopify brings generative retrieval to production search and recommendations

Starburst opens Context Studio private preview for governed agent context

OpenAI, Groupe Figaro sign ChatGPT journalism partnership
