<dataset>/<subject>.h5
    eye_block  [47, 29, 18, T]  float32   normalized BOLD around both eyes
    labels     [T, 10, 2]       float32   gaze x/y, only in labeled subsets

index.parquet at the repo root has one row per participant (dataset, subject, number of TRs, repetition time, whether labels are present, QA status), so you can select what you need before downloading anything.

Two subsets

Unlabeled — the bulk of the corpus, for self-supervised pretraining.
Labeled — participants with simultaneous eye-tracking, for fitting and evaluating a gaze probe. Same format; labels is simply present. These are the dsL##_* folders, so dsL*/*.h5 selects them without opening a file.

dsL01–dsL06 come from the DeepMReye 1.0 training sets. The rest were ingested from OpenNeuro datasets that recorded eye tracking during the scan:

folder	source	n	TR (s)	paradigm
`dsL07_deepmreye_calib`	ds006833	15	1.2	fixation / pursuit / free viewing
`dsL08_studyforrest_movie`	ds000113	15	2.0	movie (Forrest Gump), 7T
`dsL11_backtothefuture`	ds006642	37	1.5	movie (Back to the Future)

Gaze/BOLD alignment

Every ingested dataset was checked by decoding gaze from the eye block at a range of TR shifts and confirming the correlation peaks at lag 0 — the eyeball signal is not hemodynamic, so a correctly aligned recording has no delay to absorb an error. Each file records how its time origin was recovered (gaze_anchor: a BIDS StartTime, a scanner-trigger column, or a sync message in the tracker stream) and any residual offset applied (gaze_time_offset).

Gaze y grows downward — screen coordinates with a top-left origin, the same convention across every folder here. (An earlier release had dsL08, dsL09 and dsL12 inverted on this axis, and dsL12 additionally had x and y transposed. If you pulled those before 2026-08-20, re-download them. A lag sweep cannot see a sign error, which is why it took a cross-dataset readout to catch.)

Datasets that recorded gaze but are published unlabeled

Four datasets ship eye tracking, were ingested as labeled sets, and were then retired. Their eye blocks are published as ordinary unlabeled participants under their own accessions — there is no separate category for them, because what is left after removing untrustworthy labels is exactly what every other unlabeled participant is. Each failed for a different reason, and the reasons are worth distinguishing:

accession	was	why it is unlabeled
`ds001242`	`dsL09`	trigger drifts per subject and per run; 37–40% track loss; its own sidecar reports >25% data loss
`ds007532`	—	`StartTime` mixes true offsets and raw tracker clocks run by run; no dataset-level origin exists
`ds004158`	`dsL12`	labels are correct; resting state with a central fixation dot gives 0.26–1.3° gaze SD, so there is nothing to decode
`ds004283`	`dsL13`	timing is correct; the orbits are clipped in the imaging (up to 99.8% of the eye mask empty) and horizontal gaze is not recoverable

The first two are gaze-recording failures. The last two are not: ds004158's labels and ds004283's alignment are both sound, and what fails is the paradigm and the imaging respectively. A per-subject offset would have "fixed" the first two and would have been fitted on the decoding target, which is why it was not done.

Processing

Each functional run was coregistered to the DeepMReye template with ANTs (Affine, Affine, SyNAggro), masked to the eyeballs with voxels outside the mask set to 0, and cropped to a fixed [47, 29, 18] bounding box. Values are z-scored per voxel across time and per volume across space, then clipped at 5 SD. Labeled and unlabeled participants went through identical processing.

Gaze labels are sampled 10 times per TR: sub-bin j of volume t holds the mean gaze over [(t + j/10)·TR, (t + (j+1)/10)·TR), so the bins do not overlap and the mean of the ten is the mean gaze during that volume. NaN marks TRs with no valid gaze sample — mask them rather than dropping them, or the block and the gaze go out of alignment.

Units differ by dataset — read label_units from the file attributes.

Every folder here is in degrees of visual angle. The source papers were checked for each ingested dataset and a conversion is applied only where the documented geometry (display size and viewing distance) determines one; a dataset that does not document enough is stored in its native screen units with label_units saying so, rather than being given an invented conversion.

Pearson correlation is invariant to the difference. Training is not: if you fit one readout over several datasets pooled, standardise the target per dataset first, or the largest-scale dataset dominates the loss.

Loading

import h5py
from huggingface_hub import hf_hub_download

path = hf_hub_download("DeepMReye/eyeballs", "ds000001/sub-01.h5", repo_type="dataset")
with h5py.File(path) as f:
    block = f["eye_block"][..., :100]   # one 100-TR window

Files are chunked so reading a window does not decompress the whole run.

Caveat

Windows are a fixed number of TRs, not a fixed duration, and repetition times differ across datasets — so a 100-TR window is not the same amount of real time everywhere. repetition_time is in the index and in each file's attributes.

Source

Derived from OpenNeuro datasets, which carry their own licenses (typically CC0). Please cite the original datasets alongside this collection.

Downloads last month: 2

Total file size:

82.2 GB