Marraffini-Giovanni/fbert-corpus

F-BERT pretraining corpus

Parcellated fMRI windows from 162 OpenNeuro datasets (9,578 subjects, 27,835 recordings, resting state and task, about 4,000 hours of scan), the corpus on which the F-BERT-1M and F-BERT-8M encoders were pretrained, together with each recording's pretraining target. It contains derived data only: no images, no metadata beyond the dataset and recording identifiers.

The code that builds the corpus, trains the encoders and reproduces the paper is at github.com/GioMarraffini/fc-power.

Summary statistics

source datasets	162 (OpenNeuro)
subjects	9,578
recordings	27,835 (6,668 with a task name containing "rest")
windows	106,397 of 80 timepoints (median 3 per recording)
regions	450: Schaefer-400 cortex + Tian-S3 subcortex (50)
repetition time	native, 0.75 to 3 s, so an 80-timepoint window spans 1 to 4 min
preprocessing	fMRIPrep, default settings; band-pass 0.01–0.08 Hz; z-scored per region
size	26 GB

Files

file	shape	dtype	content
windows.npy	(106397, 450, 80)	float32	the windows, regions x timepoints, z-scored per region
window_to_recording.npy	(106397,)	int32	index into recordings.json for each window
teachers.npy	(27835, 101025)	float32	per recording, the pretraining target: the upper triangle of FC^0.35 of the whole recording (√2-weighted off-diagonal), centred and scaled to unit norm
recordings.json	list of 27,835 strings		<dataset>/<subject>/<subject>_task-<task> identifiers
licenses.json	dict keyed by dataset id		licence, snapshot, name, number of recordings and OpenNeuro URL of every source dataset
LICENSE_SUMMARY.json			licence counts and the list of non-commercial datasets
LICENSES.md			the same licence table, human-readable

FC^0.35 is the power-Euclidean transform of the Pearson connectome, V diag(lambda^0.35) V^T, described in the paper; the raw timeseries of a recording is the concatenation of its windows, so any other target can be recomputed from windows.npy.

Download

from huggingface_hub import snapshot_download
snapshot_download("Marraffini-Giovanni/fbert-corpus", repo_type="dataset", local_dir="release/corpus")

or, from the code repository, CORPUS=1 make fetch. Load the large arrays with numpy.load(..., mmap_mode="r").

import json, numpy as np
windows = np.load("release/corpus/windows.npy", mmap_mode="r")            # (106397, 450, 80)
teachers = np.load("release/corpus/teachers.npy", mmap_mode="r")          # (27835, 101025)
w2r = np.load("release/corpus/window_to_recording.npy")                   # (106397,)
recordings = json.load(open("release/corpus/recordings.json"))

Licensing Information

Each source dataset keeps its own OpenNeuro licence and the windows derived from it inherit that licence: 138 datasets are CC0, 6 PDDL 1.0, 16 state no licence in their dataset_description.json (OpenNeuro's upload terms require CC0), 1 is CC BY 4.0 (ds001634) and 1 is CC BY-NC 4.0 (ds001338, 66 recordings). The corpus as a whole is therefore for non-commercial use unless the windows of ds001338 are removed (its indices follow from recordings.json). licenses.json gives the licence of every dataset by id. Please cite the original datasets you use; their names and OpenNeuro pages are listed in LICENSES.md.

The five evaluation cohorts of the paper (HCP-YA, AOMIC-ID1000, ABIDE-I, ADHD-200, CoRR) are not in the corpus and are not redistributed.

Citation

A preprint is in preparation; the citation will be added here when it is available.

@article{fbert2026,
  title   = {Flattening the Connectome Spectrum: A Spectral Filter for FC Induces a Pretraining Target for fMRI Encoders},
  year    = {2026},
  note    = {Preprint to appear}
}