FUTO Builds Swipe-Typing Model on 1M Crowdsourced Gestures, Dataset Free on HuggingFace

FUTO crowdsourced its way to a new swipe-typing model by collecting real gesture data instead of relying on synthetic input. Starting in August 2024, volunteers visited a mobile webpage, consented to participation, and traced QWERTY swipes for sentences drawn mostly from Wikipedia, one word at a time. The effort yielded more than a million swipes, with a small number of low-quality samples discarded.

In March 2025, FUTO published the cleaned set of one million swipes under the permissive MIT license, hosting it on HuggingFace for anyone to use. The company leaned on this corpus both to train its own models and to benchmark competing swipe-typing approaches.

The release matters because high-quality, openly licensed gesture data is scarce, and on-device input methods typically depend on proprietary datasets. By gathering consent-based human swipes and giving the result away freely, FUTO lowers the barrier for others to build and evaluate keyboard input systems without starting from scratch.