Products
SageMaker AI Serverless Model Customization adds Full Fine-Tuning
Amazon SageMaker AI Serverless Model Customization now supports Full Fine-Tuning (FFT), allowing users to update all model parameters without provisioning GPUs or configuring distributed training. AWS compares FFT with LoRA on a Qwen3-4B model using a 110K coding instruction dataset and built-in evaluations. The post reports that FFT can outperform LoRA as dataset scale grows, while also warning that FFT is sensitive to hyperparameters and can cause catastrophic forgetting. It provides practical guidance on learning rate, epochs, batch size, and evaluation for domain-specific fine-tuning.
Amazon SageMaker AIFull Fine-TuningLoRA