From the source
Amazon says SageMaker AI Serverless Model Customization now supports Full Fine-Tuning, letting teams train a model on their own data without managing GPU clusters.
- The post contrasts FFT with parameter-efficient methods such as LoRA, which freeze the pretrained weights and inject small low-rank matrices into the model layers.
- Amazon says the low-rank structure limits the representation power of the fine-tuned model, and that as datasets scale beyond tens of thousands of examples, LoRA hits an accuracy plateau where raising rank fails to close the gap with standard fine-tuning.
- FFT instead updates the model's weight matrix uniformly, which Amazon says yields stronger convergence and better activation alignment across layers.
- In experiments on a Qwen3-4B base model fine-tuned on the Magicoder-Evol-Instruct-110K dataset of 110,000 coding instruction-response pairs, FFT reached 90.0 percent on LLM-as-Judge semantic correctness at 100K+ training examples.
- The dataset was split 90/5/5, with 99,000 records for training, 5,500 for validation and 5,500 held out for testing.






