Brain recordings from 7,272 people are rebuilt for training
A public set of brain-wave data keeps every owner's licence and still trains a model almost as well.



Picture made for ABUREAU.
Thousands of people sat still while sensors listened to their brains. Most of those recordings never help an AI model, because each set has its own format and its own licence.
A research team has now rebuilt 77 shareable brain-wave data sets into one training-ready collection of 14,075 hours from 7,272 subjects. Each data set keeps its original licence, and users can pick subsets by name.
To test it, the team trained a brain-wave model from scratch on the collection and scored 0.331 on a standard test, against 0.333 for the released model. The gap was smaller than the difference between the team's own two training runs. Fine-tuning was less close, and the authors say part of that comes from how the results were scored.
For people who own or use data, the lesson is plain. Tidy formats and clear licences, kept per data set, turn scattered records into something a buyer can use and an auditor can check. The licence travelled with the data, and that did not stop the work.

