AWS brings full fine-tuning to SageMaker AI’s serverless customization

AWS has added Full Fine-Tuning (FFT) support to SageMaker AI Serverless Model Customization, allowing businesses to update all model parameters without provisioning GPUs or configuring distributed training. The company says FFT delivered stronger results than LoRA in experiments involving a Qwen3-4B model and 110,000 coding instruction-response pairs.
- SageMaker AI Serverless Model Customization now supports Full Fine-Tuning.
- AWS compared FFT and LoRA using Qwen3-4B and 110,000 coding instruction-response pairs.
- FFT reached 90.0 percent on LLM-as-Judge semantic correctness at 100K+ training examples.
- AWS says serverless customization handles compute allocation and checkpoint management.
Why AWS is adding Full Fine-Tuning
Fine-tuning adapts a pretrained large language model to a company’s domain, task, or style by continuing its training on the company’s own data. AWS presents this as a way to improve accuracy on specialized work while potentially lowering inference costs through smaller, task-specific models.
Many customization workflows use Low-Rank Adaptation, or LoRA. This method freezes the pretrained model’s weights and adds low-rank matrices, reducing the number of trainable parameters and the GPU capacity required. AWS says that approach works well for simpler adaptations, but its low-rank structure can limit the model’s ability to represent complex changes.
FFT instead updates the model’s parameters simultaneously. According to AWS, this allows the model to use the full representational capacity of its weight matrices and better absorb complex, domain-specific knowledge. The company says LoRA can reach an accuracy plateau as tasks become more complex and datasets grow beyond tens of thousands of examples.
Benchmarking FFT against LoRA
AWS used supervised fine-tuning on the Qwen3-4B base model available in Amazon SageMaker AI Studio. The experiments used the Magicoder-Evol-Instruct-110K dataset, described by AWS as a curated collection of 110,000 coding instruction-response pairs covering algorithm design, debugging, data structures, and systems programming.
The dataset was divided into 99,000 training records, 5,500 validation records, and 5,500 held-out test records. Both LoRA and FFT received the same prompts, targets, and evaluation code.
AWS reports that FFT achieved 90.0 percent on LLM-as-Judge semantic correctness across code-generation tasks when training data reached 100K+ examples. The evaluation used Amazon Nova Pro as an automated evaluator of whether generated code was logically correct. AWS also says FFT extended beyond LoRA across every dimension in a radar-chart comparison, with particularly pronounced leads on BLEU and ROUGE-2.
The models were evaluated with LLM-as-Judge, PRIMECODE, and MMLU. These methods assessed semantic correctness, code-generation fidelity, and general knowledge retention. PRIMECODE used BLEU, ROUGE-1/2/L, and F1 to compare generated code with reference solutions.
Serverless training removes infrastructure work
SageMaker AI Serverless Model Customization is designed to let data scientists and machine learning engineers specify a base model, dataset, and hyperparameters while the service handles compute allocation and checkpoint management. AWS says the workflow requires no GPU provisioning, cluster setup, or distributed-training configuration.
The post shows an FFT job configured through the SageMaker Python SDK with a learning rate of 2e-5, a global batch size of 8, 2 maximum epochs, and a learning-rate warmup ratio of 0.1. Training metrics are tracked in the job and can be viewed in the linked MLflow app.
AWS says switching the example from FFT to LoRA requires changing only TrainingType.FULL to TrainingType.LORA. For users with large, high-quality datasets and complex customization needs, the announcement positions FFT as a way to pursue higher model quality without taking on the infrastructure management associated with full training.


