WORLDTECHAmazon makes SageMaker AI training sessions generally available for custom post-training
Amazon has made training sessions generally available as part of SageMaker AI's serverless model customization. The feature lets customers write their own post-training loop while the platform manages the GPU infrastructure beneath it.
What training sessions take off users' hands
Training sessions give customers control over post-training logic without requiring them to manage the infrastructure beneath it. Users write and coordinate the training loop themselves, including custom loss functions or reward signals. SageMaker AI takes on GPU allocation, distributed execution, fault recovery and checkpoint storage.
Because the service is serverless, users do not pick instance types, manage clusters or write distributed training configuration. Amazon presents this as a way to customize models without carrying the burden of infrastructure management.
Why teams are customizing open-weight models
According to the announcement, organisations that are getting real results from AI increasingly customize open-weight models, whose trained parameters anyone may download, rather than relying only on off-the-shelf foundation models. Post-training techniques such as reinforcement learning feature prominently in this work.
The text argues that a model trained on a company's business context, product catalogue or clinical vocabulary can be more accurate for its task and cheaper to run. For agents, the data, workflows, evaluations and success criteria also shape how the model behaves.
Multi-turn reinforcement learning and its infrastructure demands
The announcement identifies multi-turn reinforcement learning as a significant shift in agent training. Agents that handle tasks such as booking travel, sorting support tickets or carrying out multi-step research need several rounds of interaction before their outcome can be judged.
Such loops grow complicated quickly when training, inference, scoring and the next update run on separate infrastructure, with each stage needing different compute. The announcement states that SageMaker AI already offers multi-turn RL as a managed, serverless technique.
Techniques and tools on offer
Customers can apply SFT, DPO, RLVR, RLAIF or multi-turn RL. They choose how much of the model to update, through Low-Rank Adaptation (LoRA) or full fine-tuning, and can chain techniques together in continuous customization.
An agent can take a description of a use case and handle data preparation, technique choice, training, evaluation and deployment. According to the announcement, the agent works inside SageMaker Studio, Kiro, Claude Code, Codex or GitHub Copilot, and Rob Gontarek, Senior Vice President, Data Science at ImagineSoftware, is quoted in support.
Open questions
- How much of the described workflow is available beyond the techniques listed?