Amazon SageMaker AI now supports custom sequence lengths for serverless model customization jobs

Amazon SageMaker AI has introduced a new feature enabling users to set the sequence length during serverless model customization jobs. Previously, sequence lengths were fixed, limiting flexibility for tasks requiring varied input sizes. The update addresses customer requests for longer sequence lengths and greater control over fine-tuning parameters. Users can now adjust sequence lengths to better match their training data, whether processing short Q&A pairs or longer documents.

The feature supports fine-tuning for multiple open-weight model families, including gpt-oss, Gemma, Llama, Nemotron, and Qwen. Customers can now set hyperparameters such as dataset_max_len for SFT and DPO jobs, or max_prompt_length and max_response_length for RFT jobs. This allows for more precise alignment with specific use cases and data characteristics. The flexibility aims to improve model performance by accommodating diverse training scenarios.

Published maximum sequence lengths for many models now reach up to 16K tokens, depending on the model and technique used. This enhancement expands the range of tasks that can be handled, particularly those involving longer inputs. The feature is available in multiple AWS regions, including US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland). Users can access the functionality through Amazon SageMaker Studio, the SageMaker Python SDK, or coding assistant integrations.

To begin using the feature, users can navigate to the Models page in Amazon SageMaker Studio to launch a customization job. Alternatively, they can utilize the SageMaker Python SDK or leverage Agent skills with their preferred coding assistant. Detailed documentation is available to guide users through the process and provide a list of supported models and sequence lengths per technique.