Hugging Face's logo

Datasets:
windwerfer
/
neurofeed-eeg-windows

Tasks:
Other
Size:
1M<n<10M
Tags:
eeg
electroencephalography
muse
neurosity-crown
bci
time-series
License:
Dataset card Files
xet
 Community
The Dataset Viewer has been disabled on this dataset.

Neurofeed EEG Windows

Ready-to-train 2-second, 256 Hz EEG windows derived from open EEG/PSG datasets and remapped onto consumer-headset layouts: Muse (AF7, AF8, TP9, TP10) and Neurosity Crown (C3/C4-centred). Every config ships fixed subject-level splits, per-recording manifests with per-source license and upstream URL, frozen-encoder baselines, and honest negative results.

  • About 1.56 M windows in 8 configs, all from open sources (CC0-1.0, ODC-By-1.0, CC-BY-4.0).
  • Derived data only: no raw EDF/BDF, no gated or non-commercial sources, no model weights.
  • This is proxy data: PSG and research-cap electrodes mapped to headset positions, not recordings from Muse or Crown hardware.

Configs

Config Montage, X shape Labels Subjects (train/val/test) Windows Source license(s) Frozen baseline (test macro-F1) Status
muse4_vigilance_sleep_edf Muse4 proxy (N,4,512) drowsy-wake (W) / hypnagogic (N1) + stage_raw/stage_coarse 124 (86/19/19) 519,009 ODC-By-1.0 + CC-BY-4.0 CBraMod 0.747 (cross-cohort test) primary Muse vigilance release
crown2_vigilance_hmc Crown2 C3,C4 (N,2,512) W / N1 151 (106/23/22) 459,438 CC-BY-4.0 CBraMod 0.670 · REVE 0.649 Crown vigilance ship candidate
crown4_vigilance_hmc Crown4 proxy C3,C4,F6≈F4,PO4≈O2 (N,4,512) W / N1 151 (106/23/22) 459,438 CC-BY-4.0 CBraMod 0.680 · REVE 0.681 Crown vigilance ship candidate
muse4_attention_ds001787 Muse4 proxy (N,4,512) concentration / mind-wandering (thought probes) 16 (12/2/2) 5,185 CC0-1.0 LOSO 0.361 (≈ chance) research, negative result
muse4_attention_ds003969 Muse4 proxy (N,4,512) meditation / thinking block (protocol proxy) 64 (60/2/2) 51,200 CC0-1.0 LOSO 0.361 (≈ chance) research, negative result
crown8_attention_ds001787 Crown8 (N,8,512) as muse4 sibling (same windows) 16 (12/2/2) 5,185 CC0-1.0 LOSO 0.351 (≈ chance) research, negative result
crown8_attention_ds003969 Crown8 (N,8,512) as muse4 sibling (same windows) 64 (60/2/2) 51,200 CC0-1.0 LOSO 0.351 (≈ chance) research, negative result
muse4_engagement_a_eng Muse4 proxy (N,4,512) low / high engagement (5 sources) 133 persons (93/20/20) 19,706 CC0-1.0 + ODC-By-1.0 + CC-BY-4.0 CBraMod 0.548 · REVE 0.590 research, confound benchmark

Each config folder has its own card (README.md, with montage, preprocessing, label rules, splits, baselines, limitations and a load snippet) and ATTRIBUTION.md (citations, license notice, changes made).

Channel orders. Muse4 AF7, AF8, TP9, TP10 · Crown2 C3, C4 · Crown4 C3, C4, F6, PO4 · Crown8 CP3, C3, F5, PO3, PO4, F6, C4, CP4. Never mix montages in one example. Details in schemas/montages.json.

How to load

The windows are NumPy .npz files (X, y, starts, label_names, plus optional keys). datasets.load_dataset and the dataset viewer are not supported, so use huggingface_hub + NumPy:

# uv add huggingface_hub numpy
import json, numpy as np
from huggingface_hub import hf_hub_download

REPO = "windwerfer/neurofeed-eeg-windows"
path = hf_hub_download(REPO, "crown2_vigilance_hmc/windows/SN001_windows.npz", repo_type="dataset")
z = np.load(path, allow_pickle=True)   # allow_pickle only for stage_raw / stage_coarse (vigilance configs)
X, y = z["X"], z["y"]                  # X: (N, 2, 512) float32, y: (N,) int64
print([str(s) for s in z["label_names"]], X.shape)

man = json.load(open(hf_hub_download(REPO, "crown2_vigilance_hmc/windows/SN001_manifest.json", repo_type="dataset")))
print(man["subject_id"], man["license_spdx"], man["source_url"])

Each config card has a full snippet that downloads only one split (manifests → subject ids → .npz files). Pin a commit hash (revision=) when you report numbers. Schema: schemas/schema_windows.md, manifest fields: schemas/manifest_fields.md.

Scale differs by config. Vigilance configs are band-passed microvolts, attention configs are raw volts, and A-eng is z-scored. Normalise before mixing.

Splits and combining configs

  • Every config has frozen subject-level splits in <config>/splits/. No subject is in two splits within a config. Windows overlap in time, so always split by subject, never by window.
  • Do not combine muse4_vigilance_sleep_edf with crown2_vigilance_hmc / crown4_vigilance_hmc. The 24 HMC subjects of the muse4 config also appear in the crown configs, and 13 of them sit in a different split. If you must pool them, use the leak-free union split in cross_config/vigilance_hmc_leakfree.json. Results under that split are not comparable to the per-config baselines.
  • Muse4 vigilance test is cross-cohort. Test is 18 Sleep-EDF telemetry subjects (a temazepam study) + SC400, and val is mostly HMC, so the 0.747 is a cross-cohort score. The split stays frozen for comparability.
  • Attention subject IDs collide: ds001787 and ds003969 both use sub-001…, so key subjects as <dataset>/<sub>.

Baselines and negative results

All numbers use frozen encoders (CBraMod, or REVE-base where noted) with a small head on held-out subjects. Full metrics JSON (per class, per fold, confusion matrices): neurofeed_eeg_datasets/baselines.

  • Works: sleep-onset vigilance (W vs N1) transfers to headset-like channel sets: Muse4 proxy 0.747, Crown2 0.670, Crown4 0.680. REVE-base reproduces the published HMC 5-stage linear probe (balanced accuracy 0.649 vs paper 0.647).
  • Does not work (published on purpose): attention / mind-wandering from 4 or 8 channels across subjects is at chance (LOSO 0.36 CBraMod, 0.50 REVE; random ≈ 0.50, collapse ≈ 0.33–0.40). Engagement looks better than it is because of task-order confounds (0.55–0.59). Rest vs meditation (ds003816), meditation depth and EEGMAT stress vs calm smokes are also at chance; see the training-lab docs.

Notebooks and scripts

Code lives in the public training lab neurofeed_train:

Encoders are not included here. CBraMod is Apache-2.0; REVE-base is gated, so users fetch it under its own terms.

Licenses and attribution

This repository is not under a single license. Each config keeps the license(s) of its upstream source(s); see LICENSES.md and each config's ATTRIBUTION.md (citations, license notice, list of changes made). Every manifest also carries license_spdx, source_dataset and source_url.

Config Upstream SPDX
muse4_vigilance_sleep_edf PhysioNet Sleep-EDF Expanded 1.0.0; PhysioNet HMC 1.1 ODC-By-1.0; CC-BY-4.0
crown2_vigilance_hmc, crown4_vigilance_hmc PhysioNet HMC 1.1 CC-BY-4.0
muse4_attention_ds001787, crown8_attention_ds001787 OpenNeuro ds001787 1.1.1 CC0-1.0
muse4_attention_ds003969, crown8_attention_ds003969 OpenNeuro ds003969 1.0.0 CC0-1.0
muse4_engagement_a_eng OpenNeuro ds007169, ds007262, ds007554; PhysioNet EEGMAT 1.0.0; STEW (processed MONSTER mirror) CC0-1.0; ODC-By-1.0; CC-BY-4.0

Cite the upstream sources when you use these windows. The upstream authors do not endorse this derived dataset.

Not included

  • Raw EDF/BDF/PSG recordings (get them from PhysioNet / OpenNeuro under their terms).
  • Non-commercial, academic-only or gated datasets.
  • Model weights or embedding caches (including gated REVE base/positions), private caches, credentials.
  • Any claim that proxy labels are native Muse/Crown labels, clinical truth or a validated psychological state.

Changelog

  • cards-v1 (staged for review): per-config cards and ATTRIBUTION (citations, SPDX, changes made, load snippets), manifests scrubbed (portable paths, license_spdx + source_url), schema docs, cross-config leak-free vigilance split, YAML (viewer: false, data_files, size category, license_link → LICENSES.md). Window .npz files unchanged.
  • 2026-09-23: added crown2_vigilance_hmc, crown4_vigilance_hmc.
  • 2026-09-23: initial Muse4 / Crown8 release.

Packaging, schemas and split JSON: neurofeed_eeg_datasets.
Downloads last month
174
Total file size:
2.74 GB