AWS says FFT updates every model weight, where LoRA changes only small low-rank adapter matrices, and that the serverless service handles compute allocation and checkpoints without GPU provisioning or cluster setup. In experiments on Qwen3-4B trained with the 110,000-pair Magicoder-Evol-Instruct-110K coding dataset, FFT reached 90.0 percent on LLM-as-Judge semantic correctness at 100,000-plus examples, and it also led LoRA on BLEU and ROUGE-2.
The post reports that a learning rate of 2e-5 on a large dataset was safe, while rates of 5e-5 and higher caused catastrophic forgetting, with MATH scores falling to 2 to 3 percent and MMLU accuracy to 3.6 percent. Beyond three epochs, every variant tested got worse on held-out capabilities, so AWS identifies three epochs as the peak. Switching a job from FFT to LoRA takes one change to the training type in the SageMaker Python SDK.
AWS does not present FFT as uniformly better. LoRA scored 81.5 percent on LLM-as-Judge against 79.5 percent for FFT in one comparison, and both methods trailed the base model on MATH exact match.
Why it matters if you build agents
If you fine-tune on large, high-quality datasets, FFT is now available on SageMaker AI without cluster setup, so it is worth testing against LoRA on your own task. Keep the learning rate and epoch count tightly controlled, because the post shows that aggressive settings can erase general capabilities while training metrics still look healthy. Check general-knowledge benchmarks alongside task scores before choosing a method for production agents..
Get stories like this every weekday
One short email with the day's agent tooling news.