NeuralBench-EEG-Core — Cross-FM evaluation

Layer-wise linear probes on six EEG foundation models across the nine tasks with public-data YAMLs in facebookresearch/neuroai. The Core spec lists 36 EEG tasks, but only 9 ship dataset configs in the public release — the other 27 require manual access to gated corpora (TUH EEG, THINGS-images, etc.). So “NeuralBench-EEG-Core v1.0” here is the full public-data sweep: 6 FMs × 9 tasks × 10–11 probe layers × 3 seeds. Three (FM×task) cells are structurally impossible: LaBraM × {mental_arithmetic, mental_workload, motor_execution}, where the dataset has channels not in LABRAM_CHANNEL_ORDER (only fixable with InterpolatedLaBraM).

PHASE4_HEATMAP
PHASE4_RANKING

Per-FM probe depth — motor imagery / Tangermann 2012

PHASE4_DEPTHS