BCI Competition IV data set 2a (BNCI Horizon 001-2014, "Graz data set A")

Description

TODO(confirm) — the author’s catalog does not hold this dataset (§10.1 rule 3); this page carries only the §10.2 facts until a registry stub is added (§10.11 item 8).

Curriculum role: ‘Two sessions on different days were recorded for each subject. Each session is comprised of 6 runs separated by short breaks. One run consists of 48 trials (12 for each of the four possible classes), yielding a total of 288 trials per session.’ Four motor-imagery classes: left hand, right hand, both feet, tongue. Trial structure: fixation cross and a short acoustic warning tone at t = 0 s; a cue arrow at t = 2 s that stays 1.25 s; imagery held until the cross disappears at t = 6 s; then a black-screen break. Each session opens with about 5 minutes of EOG estimation in three blocks — ‘two minutes with eyes open (looking at a fixation cross on the screen), one minute with eyes closed, and one minute with eye movements’ — except A04T, where ‘due to technical problems the EOG block is shorter and contains only the eye movement condition’. Sessions are labelled T (training, labelled) and E (evaluation); the evaluation labels were released after the competition deadline.

Citation

• Paper DOI: 10.3389/fnins.2012.00055
• Dataset DOI: TODO(confirm)
• Reference: Tangermann, M., Müller, K.-R., Aertsen, A., Birbaumer, N., Braun, C., Brunner, C., Leeb, R., Mehring, C., Miller, K. J., Müller-Putz, G. R., Nolte, G., Pfurtscheller, G., Preissl, H., Schalk, G., Schlögl, A., Vidaurre, C., Waldert, S., & Blankertz, B. (2012). Review of the BCI Competition IV. Frontiers in Neuroscience 6, 55. doi:10.3389/fnins.2012.00055 — the publication the BNCI Horizon entry for 001-2014 links as its DOI, and the same DOI moabb records. Verified against Crossref 2026-09-18. The data set’s own description document is by Brunner, C., Leeb, R., Müller-Putz, G. R., Schlögl, A., & Pfurtscheller, G., ‘BCI Competition 2008 – Graz data set A’ (Institute for Knowledge Discovery, Graz University of Technology); it carries no DOI of its own.

Download and access

Field	Value
Official source	https://bnci-horizon-2020.eu/database/data-sets
Access class	open
Size	779,873,919 bytes total for the 18 .mat files on the BNCI channel (each file’s length read from its Content-Range on 2026-09-18; 37.2-46.3 MB each, mean 43.3 MB). The bbci.de GDF bundle for data set 2a is listed on its own download page as 420 MB. A notebook needs only the subjects it evaluates: one subject is both sessions, 86,578,599 bytes.
Format	gdf, mat
BIDS	false
Mirrors	BNCI Horizon 2020 database, entry ‘1. Four class motor imagery (001-2014)’ — the channel moabb downloads from. Files A01T/A01E … A09T/A09E as .mat, served from lampx.tugraz.at/~bci/database/001-2014/ (moabb 1.7.2, moabb/datasets/bnci/base.py BNCI_URL). No account, no click-through: an anonymous range request for A01T.mat returned 206 on 2026-09-18.; BCI Competition IV’s own download area at bbci.de, reached through an ‘I agree to the terms and conditions’ checkbox form. It serves ‘Data sets 2a: [ GDF files zipped (420 MB) ]’ — GDF, not .mat, so the two channels differ in file format as well as in the terms they state.

Access steps:

1. Route the site uses — moabb’s BNCI2014_001, which downloads A0{1..9}{T,E}.mat from the BNCI Horizon 2020 database with no account, no licence key and no click-through. Verified anonymously on 2026-09-18 (HTTP 206 on a range request, no cookies). The BNCI page states these data sets are ‘open access’.
2. Competition route — BCI Competition IV’s own download area at bbci.de requires ticking ‘I agree to the terms and conditions’ on a form before the download index is shown, and serves zipped GDF rather than .mat. It is a checkbox, not an account: no identity is collected and nothing is signed, so ‘open’ is recorded rather than ‘registration’. If the author reads that checkbox as a gate, access becomes ‘registration’; it changes nothing about snippets, which is already ‘no’ on the licence alone.
3. THIRD ROUTE, AND IT IS THE DEFAULT — moabb 1.7.2 fetches this dataset from NEMAR, not from BNCI Horizon 2020: BNCI2014_001.nemar_id is nm000139, a plain get_data() downloads from data.nemar.org and caches under MNE-NEMAR rather than MNE-bnci-data. The licence verified for this entry is the one BNCI Horizon 2020 publishes (CC BY-ND 4.0); NEMAR is a different publisher and its terms were NOT read, so a run that silently takes the default is fetching from a channel this site has not checked. Both Level 7 notebooks therefore set MOABB_DOWNLOAD_PROVIDER=upstream and print it. Found 2026-09-18 by notebooks-L7a. TODO(confirm): whether NEMAR republishes under the same terms.

License

• Name: CC-BY-ND-4.0
• Note (verbatim from the directory): SETTLED, and it does not qualify — this is a NO-DERIVATIVES licence and every snippet this site ships is a derivative (§10.7: ‘the site’s snippets are derivatives (re-referenced, filtered, cropped)’). Four statements, all read from primary material on 2026-09-18: one named licence, corroborated once, plus two silences. (1) The BNCI Horizon 2020 database entry for 001-2014 — the channel moabb actually downloads from, and the licensor’s own republication — states ‘License: Creative Commons Attribution No Derivatives license (CC BY-ND 4.0)’ and ‘Licensor: Institute for Knowledge Discovery, Graz University of Technology’, with the licence hyperlink pointing to creativecommons.org/licenses/by-nd/4.0/. The Institute for Knowledge Discovery is the group that recorded the data (it is the affiliation on the data set’s own description document), so this is the recording group’s own statement of terms, not a third party’s summary. The same page’s preamble reads ‘All data sets in this database are open access. This means that you can freely download and use the data according to their licenses. Additionally, if there is an associated publication, please make sure to cite it.’ (2) moabb 1.7.2 agrees: moabb/datasets/bnci/bnci_2014.py records license=“CC-BY-ND-4.0”, repository=“BNCI Horizon” for BNCI2014_001. This is a secondary source and is recorded as corroboration only; the BNCI page is the primary one. (3) SILENCE, not contradiction — the data set’s own description document (the PDF served beside the .mat files) states no licence and no terms of use at all. (4) DISAGREEING STATEMENT, on terms rather than on the licence name — BCI Competition IV’s download page at bbci.de gates its own copy behind a checkbox and states different obligations: ‘Each participant has to agree to give reference to the group(s) which recorded the data and to cite (one of) the paper listed in the respective description in each of her/his publications where one of those data sets is analyzed. Furthermore, we request each author to report any publication involving BCI Competiton data sets to us for including it in our list.’ Those are attribution and reporting obligations; they name no licence and say nothing about derivatives, so they are compatible with CC BY-ND rather than in conflict with it. No statement found anywhere permits derivatives. Per §10.7 the most restrictive reading governs and it is CC BY-ND 4.0: redistribution of the data verbatim would be permitted with attribution, but a cropped, filtered or re-referenced snippet would not, so snippets derives to ‘no’. NOTE FOR THE BUILD: derive_snippets has no explicit ND branch — it tests for ‘NC’ and for ‘SA’ by substring and returns ‘no’ for CC-BY-ND-4.0 only because the name is absent from PERMISSIVE. The answer is right; the rule is implicit. §10.7’s prose does not list ND either, because until now no directory entry was ND.
• Snippets on this site: no — the site ships no snippet or precomputed product from this dataset; lessons link the official access page.

Acquisition

Field	Value
Device	Twenty-two Ag/AgCl electrodes for EEG with 3.5 cm inter-electrode distances in a 10-20 layout, plus 3 monopolar EOG channels. ‘The sensitivity of the amplifier was set to 100 µV’ for EEG and ‘to 1 mV’ for EOG. From the dataset’s own description document (Brunner, Leeb, Müller-Putz, Schlögl & Pfurtscheller, ‘BCI Competition 2008 – Graz data set A’), read 2026-09-18 from the copy served beside the data.
Device class	research-cap
Channels	22 (22 EEG + 3 EOG, all at 250 Hz. moabb names the EEG channels Fz, FC3, FC1, FCz, FC2, FC4, C5, C3, C1, Cz, C2, C4, C6, CP3, CP1, CPz, CP2, CP4, P1, Pz, P2, POz and the EOG channels EOG1-EOG3.)
Sampling rate (Hz)	250
Online filters	Band-pass 0.5-100 Hz on both EEG and EOG, with a 50 Hz notch enabled: ‘The signals were sampled with 250 Hz and bandpass-filtered between 0.5 Hz and 100 Hz. … An additional 50 Hz notch filter was enabled to suppress line noise.’ So this data is NOT unfiltered — the description document states the analog chain and the notch.
Reference	Left mastoid reference, right mastoid ground: ‘All signals were recorded monopolarly with the left mastoid serving as reference and the right mastoid as ground.’
Mains frequency (Hz)	50
Paradigms	none
Participants	9
Population	9 subjects. The description document says only ‘This data set consists of EEG data from 9 subjects’ and records no age, sex or handedness; demographics TODO(confirm) — no primary source read here states them.
Clinical groups	none
Sessions	2
Durations	‘Two sessions on different days were recorded for each subject. Each session is comprised of 6 runs separated by short breaks. One run consists of 48 trials (12 for each of the four possible classes), yielding a total of 288 trials per session.’ Four motor-imagery classes: left hand, right hand, both feet, tongue. Trial structure: fixation cross and a short acoustic warning tone at t = 0 s; a cue arrow at t = 2 s that stays 1.25 s; imagery held until the cross disappears at t = 6 s; then a black-screen break. Each session opens with about 5 minutes of EOG estimation in three blocks — ‘two minutes with eyes open (looking at a fixation cross on the screen), one minute with eyes closed, and one minute with eye movements’ — except A04T, where ‘due to technical problems the EOG block is shorter and contains only the eye movement condition’. Sessions are labelled T (training, labelled) and E (evaluation); the evaluation labels were released after the competition deadline.

Used on this site

L7.1

Loader

Fetches only the stated subset. Recorded loader: moabb.datasets.BNCI2014_001() — moabb's own paradigm/evaluation machinery is what L7.1 uses; the underlying files come from the BNCI Horizon 2020 database.

# TODO(confirm): loader recorded from the moabb API, not from the catalog.
from moabb.datasets import BNCI2014_001

ds = BNCI2014_001()
data = ds.get_data(subjects=[1])

Caveats worth knowing

• Not in the author’s catalog: every fact here was read from the BNCI Horizon 2020 database, from the data set’s own description document and from moabb 1.7.2’s source on 2026-09-18, not from data/catalog (§10.1 rule 3). The author should add a registry stub (§10.11 item 8) carrying license.name CC-BY-ND-4.0.
• CC BY-ND forbids derivatives, so no asset derived from this dataset ships on this site — no snippet, no precomputed CSP product, no figure built from its signal. A notebook may still download it and compute on it (§10.7 class A/B download rules do not apply here: this is neither, so the lesson downloads and computes but ships nothing back). L7.1’s widget must use ds-eegbci, whose CSP products already ship.
• The data set’s own description document imposes a use restriction that is not a licence term but binds the analysis: ‘The EOG channels are provided for the subsequent application of artifact processing methods and must not be used for classification.’ A decoding lesson that leaves the 3 EOG channels in its feature set is doing the thing the dataset explicitly forbids — and would also be teaching exactly the artifact-driven decoding pf-decoding-leakage warns about.
• The description document also requires causality of any competition submission (‘All algorithms must be causal’) and asks that EOG artifacts be removed before further processing. Those are competition rules rather than licence terms, but they are the dataset authors’ stated expectations for how it is analysed.
• Not unfiltered: an analog 0.5-100 Hz band-pass and a 50 Hz notch were applied at acquisition. Anything this dataset is used to teach about line noise or about drift is being taught on already-filtered data, unlike ds-eegbci.
• Trials the experts marked as containing artifacts are flagged in the files (event type 1023, and an ArtifactSelection list), so an ‘artifact rejection’ step here can either use the experts’ marks or ignore them — which choice is made changes the accuracy, and the lesson should say which it made.
• Sessions T and E are on different days, so a within-subject train-on-T / test-on-E split is a genuine cross-session evaluation, not a random split. A random split across the pooled sessions would leak and is the pf-decoding-leakage case in miniature.
• No dataset DOI found. The DOI the BNCI entry links is the Tangermann et al. (2012) competition review, a publication rather than a data deposit, so source.dataset_doi stays TODO(confirm).

Related datasets

• ERP CORE (Compendium of Open Resources and Experiments)
• EEG Motor Movement/Imagery Dataset (EEGMMIDB)
• LEMON (MPI-Leipzig Mind-Brain-Body)
• Dortmund Vital Study resting EEG
• SRM Resting-state EEG

# Loader recorded in the catalog: moabb.datasets.BNCI2014_001() — moabb's own paradigm/evaluation machinery is what L7.1 uses; the underlying files come from the BNCI Horizon 2020 database
# MOABB fetches per-subject files on demand; restrict subjects to keep the download small.
import moabb  # noqa
# TODO(confirm) exact MOABB dataset class and subject subset