REVE

Config name: REVE

Pretrained checkpoint: brain-bzh/reve-base (HuggingFace Hub)

Reference: El Ouahidi et al., REVE: A Foundation Model for EEG – Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects, NeurIPS 2025

REVE is a vision-transformer-style EEG foundation model pretrained via masked autoencoding on over 60,000 hours of EEG data from 92 datasets spanning 25,000 subjects at 200 Hz. Its defining feature is a 4D positional encoding that jointly encodes the 3-D spatial coordinates of each electrode and the temporal position of each patch, enabling transfer across arbitrary electrode configurations without retraining.

Pretraining data overlap

Warning

REVE was pretrained on 92 public EEG datasets. Several of these overlap with NeuralBench evaluation datasets, which means results on the affected tasks may be inflated due to the model having seen the same recordings (though not the same labels) during self-supervised pretraining.

The known overlaps with NeuralBench evaluation tasks are:

Pretraining dataset	NeuralBench task	NeuralBench study	Notes
TUH (26,847 h)	pathology	Lopez2017	TUH Abnormal EEG subset
TUH	clinical_event	Harati2015	TUH EEG Events subset
TUH	artifact	Hamid2020	TUH EEG Artifact subset
Lee2019_ERP	p3	Lee2019Erp	Via MOABB
Lee2019_SSVEP	ssvep	Lee2019Ssvep	Via MOABB
PhysionetMI	motor_imagery	Schalk2009	PhysioNet eegmmidb
Schirrmeister2017	motor_execution	Schirrmeister2017	Via MOABB
Liu2024	video	Liu2024	Via MOABB
THINGS2 / ON_ds003825	image	Gifford2022Large	THINGS-EEG2 (OpenNeuro ds003825)

Known limitations

• Pretrained weights require registration — Access to the brain-bzh checkpoints on HuggingFace requires agreeing to the authors’ data usage terms.