README.md

---
license: cc-by-nc-4.0
---

This is the LibriBrain2 dataset. Together with [LibriBrain](https://huggingface.co/datasets/pnpl/LibriBrain), it forms **LibriBrain100**.

Serialised data which have been optimised for easy machine learning are included in the derivatives folders, together with linguistic annotations in event files.

Raw BIDS data are structuerd within each datset as usual. This repo contains multiple datsets:

* LibriBrain2/Sherlock1
* LibriBrain2/Sherlock8
* LibriBrain2/Sherlock9
* LibriBrain2/TIMIT
* LibriBrain2/MOCHATIMIT
* LibriBrain2/TheMoth

Please note that some of the data in Sherlock1 have been embargoed until after this year (2026)'s open machine learning competition, but will be released after. Specifically, we've included the full serialised data for sub-1 through sub-12, but reduced data for the remaining subjects. In the competition, this is being used to motivate advances in methods that generalise with decreasing amounts of data for fine-tuning, to get down to a scale of data for users that would be practical for clinical applications. To avoid data leakage during the competition, BIDS for Sherlock1 are being embargoed too.

Dataset paper to appear on arXiv:

@article{mantegna2026libribrain100,
         title={{LibriBrain100}: One Hundred Hours of Broad and Deep {MEG} Data for Neural Speech Decoding at Scale},
         author={Mantegna, Francesco and Jayalath, Dulhan and Elvers, Gereon and Kim, Tasha and Ballyk, Benjamin and Fung, Alex and Cho, SungJun and Kwon, Teyun and Kurth, Luisa and \\"Ozdogan, Miran and Landau, Gilad and Somaiya, Pratik and Voets, Natalie and Woolrich, Mark and Parker Jones, Oiwi},
         year={2026},
         journal={arXiv preprint arXiv:XXXX.XXXXX}
}