MegNIST

MegNIST is a magnetoencephalography (MEG) dataset for benchmarking non-invasive inner-speech decoding. It contains 12,000 trials from one participant imagining the digits zero through nine.

The dataset is hosted on Hugging Face at pnpl/MegNIST. Raw data are available in BIDS format, together with preprocessed derivatives and machine-learning-ready HDF5 files.

Loading the dataset

from pnpl.datasets import MegNIST

dataset = MegNIST(
    data_path="./data/MegNIST",
    partition="train",
)

x, y = dataset[0]

print(x.shape)  # (306, 250)
print(y)        # digit label 0-9

The standard partitions are:

• train: 10,000 trials

• validation / val: 1,000 trials

• test: 1,000 trials

Each sample contains 306 MEG channels and 250 time samples at 250 Hz.

Task

The default task is 10-class imagined-digit classification:

from pnpl.datasets import MegNIST
from pnpl.tasks import DigitClassification

dataset = MegNIST(
    data_path="./data/MegNIST",
    task=DigitClassification(),
    partition="train",
)

Labels map directly onto the imagined digits:

Label	Digit
0	zero
1	one
2	two
3	three
4	four
5	five
6	six
7	seven
8	eight
9	nine

Standardisation

By default, MegNIST returns the released preprocessed epochs without additional normalisation (standardize=False).

PNPL can optionally apply channel-wise standardisation:

train = MegNIST(
    data_path="./data/MegNIST",
    partition="train",
    standardize=True,
)

# Reuse the train statistics for validation/test instead of fitting new ones
val = MegNIST(
    data_path="./data/MegNIST",
    partition="validation",
    standardize=True,
    channel_means=train.channel_means,
    channel_stds=train.channel_stds,
)

With standardize=True and no explicit channel_means/channel_stds, the statistics are computed from the partition being loaded. Pass the train statistics to the validation and test partitions, as above, to avoid fitting normalisation on held-out data.

Channel-wise standardisation estimates one mean and standard deviation per MEG sensor, rather than separately for every time point. It is PNPL’s default for continuous/streaming datasets, where time points do not have a fixed event-locked meaning and feature-wise standardisation would depend on an arbitrary position within each window.

For fixed, event-locked epochs such as MegNIST, other choices are possible. For example, the MegNIST baseline uses train-fitted feature-wise scaling. If you need a different normalisation scheme, leave standardize=False and apply it separately in your analysis pipeline.

Data source

Hugging Face dataset: pnpl/MegNIST

Alongside the serialised HDF5 splits, the repository contains the raw BIDS recordings (4 sessions × 3 runs), the intermediate preprocessing derivatives (bads+headpos+sss+notch+bp+ds), head-position files and preprocessing reports.

Licence and citation

MegNIST is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) licence. It is free for research and other non-commercial use; for commercial licensing enquiries contact Oxford University Innovation.

If you use MegNIST, please cite:

Kwon, T., Somaiya, P., Elvers, G., Mantegna, F., Camera, A., Braeutigam, S., Woolrich, M. and Parker Jones, O. (2026). MegNIST: A Benchmark for Non-Invasive Inner Speech Decoding.

@misc{kwon2026megnist,
  title={{MegNIST}: A Benchmark for Non-Invasive Inner Speech Decoding},
  author={Kwon, Teyun and Somaiya, Pratik and Elvers, Gereon and Mantegna, Francesco and Camera, Anna and Braeutigam, Sven and Woolrich, Mark and Parker Jones, Oiwi},
  year={2026}
}