Skip to content

SageMaker adds full fine-tuning to its serverless customization

Short answer

Amazon SageMaker AI Serverless Model Customization now supports full fine-tuning, which updates every model weight. AWS's own tests on a code task show it beating LoRA once training data passes 100K examples. The same tests show it can erase general knowledge if learning rate or epochs are set too high.

What this means for operators

Most 10-200 person B2B companies will not need this: if your support or sales assistant works from a few thousand examples, LoRA remains the sensible default, and the source reports no advantage for full fine-tuning at that scale. It becomes relevant if you hold a large, clean archive of tickets, call summaries or structured outputs, and you want a small specialised model to replace a pricier general one. Before committing, run both methods on your own held-out set, since AWS's tests show a one-line switch between them. Also add a general-knowledge check, because a model that gets better at your task can quietly get worse at everything else.

AWS has added full fine-tuning (FFT) to Amazon SageMaker AI Serverless Model Customization, according to a post on the AWS Machine Learning Blog. The service already handled LoRA, a lighter method that trains small add-on matrices. FFT instead updates all of the model's weights. AWS says users do not provision GPUs, set up clusters or configure distributed training. The post does not give pricing or a rollout date.

What AWS tested

The experiments used supervised fine-tuning on the Qwen3-4B base model with the Magicoder-Evol-Instruct-110K dataset of 110,000 coding instruction pairs. AWS split it into 99,000 training records, 5,500 for validation and 5,500 held out for testing. Both methods received identical prompts, targets and evaluation code. The post adds that the team also ran experiments on DeepSeek-R1-1.5B, Nemotron-30B and Gemma-4-26B. Each run took a few hours from submission to results, and finished models were registered automatically in SageMaker Model Registry.

At 100K+ training examples, FFT reached 90.0 percent on an LLM-as-Judge correctness score for code generation. AWS reports the gap over LoRA widens as the dataset grows, with LoRA hitting a plateau on complex tasks that more trainable parameters do not fix. These are AWS's own results on one coding task, so they say little about customer-support or sales text.

Where it goes wrong

The post is open about the downsides:

  • FFT is not always better. In one comparison LoRA scored 81.5 percent on LLM-as-Judge against 79.5 percent for FFT.
  • Both methods fell below the base model on MATH exact match.
  • Learning rates of 5e-5 or higher were destructive. MATH scores dropped to 2-3 percent and MMLU accuracy to 3.6 percent, which AWS calls catastrophic forgetting.
  • Going beyond three epochs degraded every variant tested, even though training loss kept improving.

The configuration that best preserved general knowledge, measured on MMLU, used a 5 percent warmup, batch size 8 and a learning rate of 2e-5. AWS presents these as starting points, since good values vary by task.

Getting started

The SageMaker Python SDK launches the job. Switching between FFT and LoRA means changing one training-type setting in the trainer, and the rest of the code stays the same. Built-in evaluation covers LLM-as-Judge, a code-similarity scorer called PRIMECODE, and MMLU as a check on forgotten general knowledge. The post points to AWS documentation for the list of base models that support FFT.

Source: AWS Machine Learning Blog

Next step

Visibility Analyzer

An AI news item will not tell you how assistants see your own site. The Visibility Analyzer checks that on your live site: which answers cite you, which pages an engine cannot retrieve, and what to fix first. Free to run.

Run a free visibility audit

Free to run. No card.

Fee
Free
Length
One run, minutes

Free tier: two analyses a day, no card required.