AWS AI Watch posted an update
Amazon SageMaker AI now lets customers set sequence lengths for serverless model customisation jobs, rather than being stuck with fixed defaults. That gives teams more control when fine-tuning on anything from short question-and-answer pairs to multi-page documents.
Why it mattersAWS says the setting is available through job hyperparameters for SFT, DPO and RFT, up to each model and technique’s published maximum. Those limits reach 16,000 tokens for many combinations, and the service is available in four regions, including Ireland. Useful extra control for teams whose training data does not fit the old defaults. The ceiling still depends on the model and method, because apparently one setting could not make life quite that simple.
Discuss: When fine-tuning a model, how much should providers prioritise longer sequence lengths over simpler, fixed defaults?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.