LUNA

Config name: NtLuna

Pretrained checkpoint: thorir/LUNA (HuggingFace Hub, manual download)

Reference: Döner et al., LUNA: Efficient and Topology-Agnostic Foundation Model for EEG Signal Analysis, NeurIPS 2025

LUNA is a topology-invariant EEG foundation model that processes signals from varying numbers of channels using a learned channel-unification mechanism based on cross-attention queries. It is topology-agnostic: it accepts arbitrary montages without requiring a fixed channel set. Three pretrained checkpoints are available (Base, Large, Huge); the NeuralBench default uses the Large variant. The model was pretrained on the Temple University Hospital EEG corpus (TUEG) plus the Siena Scalp EEG Database at 256 Hz.

Pretraining data overlap

Warning

LUNA was pretrained on TUEG plus the Siena Scalp EEG Database (Section 4.1 of the paper). The paper explicitly excluded TUAB, TUAR, TUSL, and SEED-V from pretraining, but did not exclude TUEV. This means clinical_event has a direct overlap, while pathology and artifact have corpus-level overlap (same TUEG parent corpus, but the specific recordings were excluded).

Pretraining dataset	NeuralBench task	Overlap type	NeuralBench study
TUEV (not excluded)	clinical_event	direct	Harati2015
TUEG (TUAB excluded)	pathology	corpus	Lopez2017
TUEG (TUAR excluded)	artifact	corpus	Hamid2020

Known limitations

• Patch size padding – n_times is automatically zero-padded to the next multiple of patch_size (40 at 256 Hz) when needed. This is handled transparently by _LunaEncoderWrapper but may introduce a small number of padding tokens at the end of the sequence.