Datasets:

Abdulkader
/

eeg-foundation-leaderboard

id stringlengths 3 15	name stringlengths 3 15	year int64 2.02k 2.03k	venue stringlengths 8 28	architecture stringclasses 4 values	pretrain_objective stringclasses 6 values	params_m float64 0 1.7k	pretrain_hours float64 43 40k ⌀	pretrain_corpora stringlengths 13 50	paper_url stringlengths 32 46	code_url stringlengths 0 41	weights_url stringclasses 1 value	license stringclasses 4 values	is_foundation_model bool 2 classes	notes stringlengths 0 234
labram-huge	LaBraM-Huge	2,024	ICLR 2024	Transformer	Masked	369	2,500	~2,500 h across >20 public EEG datasets	https://arxiv.org/abs/2405.18765	https://github.com/935963004/LaBraM		Apache-2.0	true	Neural tokenizer + vector-quantised masked prediction.
labram-large	LaBraM-Large	2,024	ICLR 2024	Transformer	Masked	46	2,500	~2,500 h across >20 public EEG datasets	https://arxiv.org/abs/2405.18765	https://github.com/935963004/LaBraM	null	Apache-2.0	true
labram-base	LaBraM-Base	2,024	ICLR 2024	Transformer	Masked	5.8	2,500	~2,500 h across >20 public EEG datasets	https://arxiv.org/abs/2405.18765	https://github.com/935963004/LaBraM	null	Apache-2.0	true
biot	BIOT	2,023	NeurIPS 2023	Transformer	Masked	3.3	5,000	SHHS + PREST + TUAB (cross-data pretraining)	https://arxiv.org/abs/2305.10351	https://github.com/ycq091044/BIOT	null	MIT	true	Biosignal transformer with channel/position embeddings for variable montages.
eegpt	EEGPT	2,024	NeurIPS 2024	Transformer	Masked	101	2,000	~2,000 h, mixed public corpora	https://arxiv.org/abs/2407.07764	https://github.com/BINE022/EEGPT	null	MIT	true	Dual self-supervised (mask-reconstruct + alignment) with spatio-temporal patching.
cbramod	CBraMod	2,025	ICLR 2025	Transformer	Masked	4.9	27,000	TUEG (~27k h)	https://arxiv.org/abs/2412.07236	https://github.com/wjq-learning/CBraMod	null	MIT	true	Criss-cross transformer with asymmetric conditional positional encoding.
neurolm	NeuroLM	2,025	ICLR 2025	Transformer	Autoregressive	1,700	25,000	~25,000 h EEG, text-aligned via GPT-2 backbone	https://arxiv.org/abs/2409.00101	https://github.com/935963004/NeuroLM	null	Apache-2.0	true	Multi-task EEG+language model; instruction-tuned across benchmarks.
bendr	BENDR	2,021	Front. Hum. Neurosci. 2021	CNN-Transformer	Contrastive	7.3	1,500	TUEG (~1.5 TB raw)	https://arxiv.org/abs/2007.14511	https://github.com/SPOClab-ca/BENDR	null	BSD-3-Clause	true	wav2vec-style contrastive pretraining adapted to EEG.
brainbert	BrainBERT	2,023	ICLR 2023	Transformer	Masked	43	43	intracranial sEEG, 43 h / 10 subjects	https://arxiv.org/abs/2302.14367	https://github.com/czlwang/BrainBERT	null	MIT	true	Pretrained on sEEG spectrograms; scalp-EEG transfer is out-of-domain.
neuro-gpt	Neuro-GPT	2,024	ISBI 2024	CNN-Transformer	Autoregressive	79.5	20,000	TUEG (~20k recordings)	https://arxiv.org/abs/2311.03764	https://github.com/wenhui0206/NeuroGPT	null	MIT	true	EEG encoder + GPT-2 predicting masked future chunks.
eeg2rep	EEG2Rep	2,024	KDD 2024	Transformer	JEPA	4.1	null	TUEG subset + per-dataset SSL	https://arxiv.org/abs/2402.17772	https://github.com/Navidfoumani/EEG2Rep	null	MIT	true	Joint-embedding predictive architecture in latent space (no pixel/signal reconstruction).
brainwave	BrainWave	2,024	arXiv 2024	Transformer	Contrastive	157	40,000	~40,000 h combined EEG + iEEG	https://arxiv.org/abs/2402.10251		null		true	Joint scalp + intracranial pretraining.
eegformer	EEGFormer	2,024	arXiv 2024	Transformer	Masked	6.5	1,100	TUEG (~1.1 TB)	https://arxiv.org/abs/2401.10278		null		true	Vector-quantised discrete representation learning.
mmm	MMM	2,024	ICML 2024	Transformer	Masked	40	null	TUEG + MI corpora; montage-agnostic topology	https://arxiv.org/abs/2407.15362		null		true	Multi-montage, multi-scale masked modelling with unified topology tokens.
sparcnet	SPaRCNet	2,023	Neurology 2023	CNN	Supervised	0.79	null	— trained from scratch	https://doi.org/10.1212/WNL.0000000000207127	https://github.com/bdsp-core/IIIC	null		false	Dense-net style CNN; the clinical IIIC baseline.
contrawr	ContraWR	2,022	J. Biomed. Inform. 2022	CNN	Contrastive	1.6	null	self-supervised on target corpus	https://arxiv.org/abs/2110.15278	https://github.com/ycq091044/ContraWR	null	MIT	false	Contrastive learning with a world-representation negative.
cnn-transformer	CNN-Transformer	2,022	EMBC 2022	CNN-Transformer	Supervised	3.2	null	— trained from scratch	https://doi.org/10.1109/EMBC48229.2022.9871576		null		false	Standard supervised baseline in the BIOT/LaBraM suites.
ffcl	FFCL	2,022	Biomed. Signal Process. 2022	Hybrid	Supervised	2.4	null	— trained from scratch	https://doi.org/10.1016/j.bspc.2022.103833		null		false	Feature-fusion CNN+LSTM baseline.
st-transformer	ST-Transformer	2,022	arXiv 2022	Transformer	Supervised	3.5	null	— trained from scratch	https://arxiv.org/abs/2206.10897		null		false	Spatio-temporal attention baseline.
eegnet	EEGNet	2,018	J. Neural Eng. 2018	CNN	Supervised	0.003	null	— trained from scratch	https://arxiv.org/abs/1611.08024	https://github.com/vlawhern/arl-eegmodels	null	Apache-2.0	false	The compact-CNN reference baseline for BCI tasks.
eegconformer	EEG-Conformer	2,023	IEEE TNSRE 2023	CNN-Transformer	Supervised	0.8	null	— trained from scratch	https://doi.org/10.1109/TNSRE.2022.3230250	https://github.com/eeyhsong/EEG-Conformer	null	MIT	false	Strong supervised MI baseline; convolutional + self-attention.
diffeeg	DiffEEG	2,026	arXiv 2026	CNN-Transformer	Diffusion	9.6	null	1.3M unlabeled segments, TUH Seizure Corpus (TUSZ)	https://arxiv.org/abs/2607.11578				true	1D U-Net with FiLM conditioning and multi-head self-attention; denoising diffusion pretraining plus RL-based fine-tuning. 9.6M params full, 6.5M encoder-only used downstream. Partial fine-tuning: early conv/downsampling blocks frozen.

EEG Foundation Model Leaderboard — data

Backing data for the EEG Foundation Model Leaderboard Space.

Files

models.json — one record per model: architecture, pretraining objective, parameter count, pretraining corpus scale, paper/code/weights links.
results.json — one record per (model, benchmark, protocol), with a metrics dict holding every metric that benchmark reports.

Loading

from datasets import load_dataset

models  = load_dataset("Abdulkader/eeg-foundation-leaderboard", data_files="models.json",  split="train")
results = load_dataset("Abdulkader/eeg-foundation-leaderboard", data_files="results.json", split="train")

Or read the raw JSON with no datasets dependency:

import json, urllib.request
url = "https://huggingface.co/datasets/Abdulkader/eeg-foundation-leaderboard/resolve/main/results.json"
results = json.loads(urllib.request.urlopen(url).read())

Result schema

{
  "model_id": "labram-huge",
  "benchmark_id": "TUEV",
  "protocol": "finetune",
  "metrics": {"balanced_acc": 0.6616, "cohen_kappa": 0.6745, "weighted_f1": 0.8329},
  "std": {"balanced_acc": 0.0170},
  "n_seeds": 5,
  "source": "LaBraM (ICLR 2024), Table 3",
  "source_url": "https://arxiv.org/abs/2405.18765",
  "verified": false,
  "notes": ""
}

protocol is one of finetune, linear_probe, frozen_knn, zero_shot, scratch. Metric values are fractions in [0, 1] (or [-1, 1] for κ and MCC), never percentages. fpr_per_hour is a count per hour, so it is not bounded.

⚠️ Verification status

verified: false means no maintainer has checked that number against the cited source yet. The initial seed covers TUAB and TUEV only, transcribed from the LaBraM (ICLR 2024) comparison tables, and is entirely unverified. Treat unverified rows as leads, not citations.

Contributing

Open a PR against this repo, or use the Submit tab on the Space, which validates the payload and opens the PR for you.

Downloads last month: 117

Number of rows:

Total file size:

26.2 kB

Space using Abdulkader/eeg-foundation-leaderboard 1

Paper for Abdulkader/eeg-foundation-leaderboard

Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI

Paper • 2405.18765 • Published May 29, 2024