README.md

---
license: cc-by-nd-4.0
pretty_name: BCI Competition IV Dataset 2a for AGTCNet
size_categories:
  - 100MB<n<1GB
tags:
  - neuroscience
  - electroencephalography
  - brain-computer-interface
  - motor-imagery
  - matlab
  - bci-competition
---

# BCI Competition IV Dataset 2a

This is an **unofficial mirror** of the four-class motor-imagery dataset listed
by BNCI Horizon 2020 as dataset `001-2014`, originally released as BCI
Competition IV Dataset 2a. It is provided for reproducing Lim et al.'s AGTCNet
experiments and is not affiliated with or endorsed by the dataset contributors,
their institutions, BNCI Horizon 2020, the competition organizers, or the
AGTCNet authors.

## Source and documentation

- Original dataset: [BNCI Horizon 2020, 001-2014](https://bnci-horizon-2020.eu/database/data-sets)
- Original competition: [BCI Competition IV](https://www.bbci.de/competition/iv/)
- Dataset description: [`description.pdf`](./description.pdf)
- AGTCNet article: [Lim et al., *IEEE Access* (2025)](https://doi.org/10.1109/ACCESS.2025.3627419)
- AGTCNet code: [`galvinlim/AGTCNet`](https://github.com/galvinlim/AGTCNet)
- Mirror-generated checksums: [`MIRROR_SHA256SUMS.txt`](./MIRROR_SHA256SUMS.txt)

The 18 scientific MATLAB files and the original description PDF are unchanged
from the BNCI Horizon downloads. This dataset card and the checksum manifest are
the only added files. The manifest was generated after download and is not an
upstream publisher manifest.

The AGTCNet paper also evaluates PhysioNet EEGMMIDB. That dataset is available
in the separate [`raei/Schalk-2004-BCI2000GeneralPurposeBrainComputerInterfaceSystem`](https://huggingface.co/datasets/raei/Schalk-2004-BCI2000GeneralPurposeBrainComputerInterfaceSystem)
mirror and is intentionally not duplicated here.

## Dataset summary

The dataset contains two sessions recorded on different days from each of nine
participants. Every session has 288 cue-based trials, balanced across four
motor-imagery classes:

1. Left hand
2. Right hand
3. Both feet
4. Tongue

Signals comprise 22 EEG and three EOG channels sampled at 250 Hz. They were
band-pass filtered from 0.5 to 100 Hz during acquisition, with a 50 Hz notch
filter. EOG channels are included for artifact processing and must not be used
for classification under the original competition protocol.

Files named `A01T.mat` through `A09T.mat` are training sessions; files named
`A01E.mat` through `A09E.mat` are evaluation sessions. These post-competition
MATLAB distributions contain labels for both session types and are the format
used by the released AGTCNet loader. Each file contains 288 labels from classes
1 through 4. `A04T.mat` has a shorter initial EOG block because of a documented
technical problem.

## Human-participant data

These are human EEG recordings. Users are responsible for following applicable
institutional and legal requirements and should not attempt to identify
participants.

## License

BNCI Horizon 2020 lists the dataset under the [Creative Commons
Attribution-NoDerivatives 4.0 International](https://creativecommons.org/licenses/by-nd/4.0/)
license, with the Institute for Knowledge Discovery at Graz University of
Technology as licensor. The license permits sharing unchanged copies with
appropriate attribution and license notice; modified versions may not be
distributed.

## Citation

Please cite the dataset description authors and the BCI Competition IV review:

> Brunner, C., Leeb, R., Müller-Putz, G. R., Schlögl, A., & Pfurtscheller, G.
> (2008). BCI Competition 2008 — Graz data set A.

> Tangermann, M., Müller, K.-R., Aertsen, A., et al. (2012). Review of the BCI
> Competition IV. *Frontiers in Neuroscience*, 6, 55.
> https://doi.org/10.3389/fnins.2012.00055