What changed
Amazon SageMaker AI now lets serverless model-customization jobs set dataset_max_len for SFT and DPO or max_prompt_length and max_response_length for RFT, up to each model and technique's published maximum. AWS lists support across gpt-oss, Gemma, Llama, Nemotron and Qwen families.
Why it matters
Teams can align training windows with their data instead of accepting fixed recipe defaults. Longer sequences can preserve multi-page examples, but they also change memory, throughput, cost and truncation behavior.
What to do
Profile the real token-length distribution, choose the smallest length that covers the target percentile, and compare loss, task quality, truncation, job time and cost against the current recipe before promoting a tuned model.
Caveat
Published maximums vary by model and technique and are only up to 16K for many models. Longer inputs do not guarantee better results and can increase cost; the release is limited to the listed SageMaker regions.