Understand this story
The English summary is being updated. See the reporting timeline below.
Latest development · Report
AWS Launches Full Fine-Tuning Support for SageMaker AI Serverless Model CustomizationDevelopments
1 confirmed developments
- Source:AWS Machine Learning
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 19
Unite.AI
AWS Launches Full Fine-Tuning Support for SageMaker AI Serverless Model Customization
On November 18, 2026, AWS launched full fine-tuning support for SageMaker AI Serverless Model Customization, eliminating the need for users to configure GPUs, clusters, or distributed training environments. In experiments using Qwen3-4B and more than 100,000 training samples, AWS achieved 90.0% semantic accuracy, although in another LLM-as-Judge comparison, LoRA scored 81.5%, outperforming FFT at 79.5%. Tests also showed that at learning rates of 5e-5 or higher, the MATH score fell to 2–3% and MMLU accuracy dropped to 3.6%, indicating that aggressive settings can cause severe catastrophic forgetting.
AWS Machine LearningFirst-party
Full Fine-Tuning Methods and Evaluation with SageMaker AI Serverless Model Customization
AWS explains how to perform full fine-tuning of Qwen3-4B with SageMaker AI Serverless Model Customization, comparing it with LoRA in terms of code generation quality and retention of general knowledge. The experiment uses 110,000 coding instruction-response pairs. With more than 100K training samples, FFT achieves an LLM-as-Judge semantic correctness rate of 90.0%. FFT is sensitive to hyperparameters: when LR ≥ 5e-5, MATH falls to 2–3 percent and MMLU to 3.6 percent; under some configurations, LoRA also scores higher than FFT in LLM-as-Judge evaluations.
AI Trends