Full Fine-Tuning (FFT) on SageMaker AI Serverless Model Customization
Full Fine-Tuning (FFT) on SageMaker AI Serverless Model Customization
Amazon SageMaker AI Serverless Model Customization now supports Full Fine-Tuning (FFT), which updates all model parameters to overcome the low-rank limits of LoRA and requires no GPU provisioning or cluster setup. In experiments fine-tuning Qwen3-4B on the 110K-example Magicoder-Evol-Instruct-110K code dataset, FFT reached 90.0 percent LLM-as-Judge semantic correctness at 100K+ training examples, though results were highly sensitive to hyperparameters, with aggressive learning rates (LR ≥ 5e-5) causing catastrophic forgetting. The post also notes LoRA slightly outperformed FFT on LLM-as-Judge (81.5 vs. 79.5 percent) and that both underperformed the base model on MATH exact match.
Source: AWS Machine Learning Blog · aws.amazon.com