Article URL: https://swipe.futo.tech/ Comments URL: https://news.ycombinator.com/item?id=48648619 Points: 105 # Comments: 35

In August 2024, we launched a dataset collection effort on the swipe.futo.org domain to collect QWERTY English swipes. Users would voluntarily visit the webpage on their mobile phone and be given instructions and information about the dataset. After consenting, they would be given sentences, primarily from Wikipedia, and would be asked to swipe them word-by-word. In the end, this produced over 1 million swipes. We filtered out a small set of low-quality swipes. In March 2025, we released a dataset of 1 million swipes under the MIT license, and it is available today on HuggingFace. We made heavy use of this data to train our models and to evaluate different swipe typing systems. The Encoder model is a universal layout-agnostic and language-agnostic, and is used for making swipe typing predictions in the general case. However, it does not offer cutting-edge accuracy. The ContextLM model is a very small language model that is trained for a single language. It's used to improve the quality of predictions by eliminating nonsensical words given the preceding words in the sentence. It only requires text data for training. Finally, the decoder is a language-specific and layout-specific model that learns layout's peculiarities and achieves leading accuracy. As it requires swipe typing data for a specific layout and language for training, we only have a QWERTY English decoder for now. With all 3 models and with a beam width of 300, we achieve a top-4 fail rate of only ~4% on our test set. Ignoring out-of-vocabulary cases, the error rate is below 1%. The encoder model is just 635,140 parameters, and the decoder is 304,155 extra. The biggest one is the ContextLM at 1.5 million, but 1.1 million of that is just embeddings. This brings us to 1,364,271 active parameters, or 2,494,767 total parameters.