# Small real datasets for preview testing Verified 2026-08-14 against the repositories' own pages, APIs, object metadata, and checksum manifests. These are public research data, not generated fixtures. Prefer the version-pinned URLs below over a repository's moving `draft` or unversioned object URL. ## Shortlist | Format | Minimal test data | Download size | License | Works in the current plugin? | | --- | --- | ---: | --- | --- | | NIfTI-1 in BIDS | OpenNeuro `ds000005`, `sub-01_T1w.nii.gz` | 5,515,017 B compressed; 11,796,832 B as `.nii` | PDDL 1.0; attribution/community norms requested | **Yes, after external gzip decompression.** The plugin accepts the resulting single-file NIfTI-1 `.nii`, but not `.nii.gz` or BIDS as a dataset. | | EDF+ EEG | PhysioNet EEG Motor Movement/Imagery, `S001R01.edf` | 1,275,936 B (1.22 MiB) | ODC Attribution 1.0 | **No.** EDF/EDF+ is not implemented. | | BrainVision EEG | OpenNeuro `ds007629`, subject 10014, ReMind run 1 | 43,190,543 B (41.19 MiB) for `.eeg` + `.vhdr` + `.vmrk` | CC0 1.0 | **No.** BrainVision and linked multi-file inputs are not implemented. | | NWB | DANDI `000006`, mouse session `20170316` | 609,568 B (0.58 MiB) | CC BY 4.0 | **No.** NWB is not implemented; the roadmap proposes an optional Python worker. | The default host file limit is 256 MiB, so every individual file here is below that limit. Format support, rather than size, blocks the last three rows. ## 1. NIfTI/BIDS: OpenNeuro mixed-gambles T1w This is a real human T1-weighted structural scan from the OpenfMRI/OpenNeuro mixed-gambles study (`ds000005`). It is a useful happy-path test because gzip decompression produces a valid single-file NIfTI-1 image: `sizeof_hdr` is 348 and magic is `n+1`. - Version-pinned download: - Compressed size: **5,515,017 B** (5.26 MiB). - Decompressed size: **11,796,832 B** (11.25 MiB). - SHA-256, downloaded `.nii.gz`: `75c615d7cce397be7a967ba31c932872488ff514586adf302cf8f73c7b4ff6b7`. - SHA-256, decompressed `.nii`: `91047a7abb6a455fe6bf9c597a573c453f7d62201bfca0cfbfd4f6578192d6ac`. - Dataset metadata and terms: . The embedded license is the Open Data Commons Public Domain Dedication and License 1.0, with attribution/share-alike community norms requested. - Requested citation from that metadata: Tom, S. M.; Fox, C. R.; Trepel, C.; Poldrack, R. A. (2007), “The neural basis of loss aversion in decision-making under risk,” *Science* 315(5811):515–518. Reproducible preparation: ```bash curl -L --fail \ 'https://s3.amazonaws.com/openneuro.org/ds000005/sub-01/anat/sub-01_T1w.nii.gz?versionId=N.vvdRaCvbawrtFePSoW51oshmP3ASq1' \ -o sub-01_T1w.nii.gz printf '%s %s\n' \ 75c615d7cce397be7a967ba31c932872488ff514586adf302cf8f73c7b4ff6b7 \ sub-01_T1w.nii.gz | sha256sum -c - gzip -dk sub-01_T1w.nii.gz printf '%s %s\n' \ 91047a7abb6a455fe6bf9c597a573c453f7d62201bfca0cfbfd4f6578192d6ac \ sub-01_T1w.nii | sha256sum -c - ``` Privacy: this is a human structural image, so treat it as sensitive even though it is public. OpenNeuro requires structural scans to be defaced **or** to have explicit participant consent and ethical authorization for publication; that policy is not proof that this particular file is anonymous. Do not attempt re-identification. See OpenNeuro's official [upload rules](https://docs.openneuro.org/user_guide.html#uploading-your-dataset). Current-plugin note: pass `sub-01_T1w.nii`, not the `.nii.gz`. BIDS sidecars and directory semantics are ignored, and anatomical reorientation is not yet implemented, so the preview follows voxel storage order. ## 2. EDF+: PhysioNet EEG Motor Movement/Imagery `S001R01.edf` is a one-minute, 64-channel baseline recording from a dataset of 109 human volunteers. It is small enough for a fast parser/metadata/signal preview fixture and contains EDF+ annotations. - Direct versioned download: . - Size: **1,275,936 B** (1.22 MiB). - SHA-256: `4743b736131a7e147c150e8b37711029b6cda5e356c4b3e8261a03cdcaaf8b0c`. - First-party checksum manifest: . - Official dataset page: . - License: [Open Data Commons Attribution License 1.0](https://physionet.org/content/eegmmidb/view-license/1.0.0/); retain attribution and notices required there. - Dataset citation: Schalk, G. (2009), *EEG Motor Movement/Imagery Dataset* (version 1.0.0), PhysioNet, RRID:SCR_007345, . The landing page also requests the standard PhysioNet citation and cites the BCI2000 paper. Verification: ```bash curl -L --fail \ https://physionet.org/files/eegmmidb/1.0.0/S001/S001R01.edf \ -o S001R01.edf curl -L --fail \ https://physionet.org/files/eegmmidb/1.0.0/SHA256SUMS.txt \ -o SHA256SUMS.txt grep ' S001/S001R01.edf$' SHA256SUMS.txt | sed 's# S001/# #' | sha256sum -c - ``` Privacy: this is human-volunteer EEG. The public EDF header's patient field is masked (`X X X X`) and the filename uses a study subject code, but the dataset page does not make a blanket anonymity claim. Treat it as de-identified or pseudonymized research data, not as proof of irreversible anonymization, and do not attempt re-identification. Current-plugin note: expected to fail as unsupported today; use this file to drive a future EDF/EDF+ adapter and annotation-channel tests. ## 3. BrainVision: OpenNeuro ROAMM run 1 ROAMM (`ds007629`, snapshot 1.3.0) is real human EEG collected during naturalistic reading. BrainVision is a three-file format, so all three files must keep their original basenames in the same directory; the `.vhdr` links to the binary `.eeg` and marker `.vmrk` by relative name. | File | Bytes | SHA-256 | Version-pinned download | | --- | ---: | --- | --- | | `sub-10014_task-ReMind_run-1_eeg.eeg` | 43,188,224 | `319bf1442414e8f57427916337ff0b59bfb1c2c1ae4b1857cfd110e7e6987dce` | [S3 object](https://s3.amazonaws.com/openneuro.org/ds007629/sub-10014/eeg/sub-10014_task-ReMind_run-1_eeg.eeg?versionId=PMzdOcbZJeYAazYpJishdAqUF90qIJfG) | | `sub-10014_task-ReMind_run-1_eeg.vhdr` | 1,801 | `f5e8699399ffac35845141a8b701fa6eb763034a2f9378f9989dfb7b8f8740e4` | [S3 object](https://s3.amazonaws.com/openneuro.org/ds007629/sub-10014/eeg/sub-10014_task-ReMind_run-1_eeg.vhdr?versionId=ofODexBmh3wxDrFBjwG7pf1ATJfpfcAu) | | `sub-10014_task-ReMind_run-1_eeg.vmrk` | 518 | `780a7bc5576c3e3e5dad4dbc642a347c176a3ba3dd3b5b26bedf0771c3084dd6` | [S3 object](https://s3.amazonaws.com/openneuro.org/ds007629/sub-10014/eeg/sub-10014_task-ReMind_run-1_eeg.vmrk?versionId=n443XwZo_9n_uRwqPXmWhgwayn4.4JqM) | - Snapshot landing page and version DOI: and . - Versioned dataset metadata: . - License: CC0 1.0, as declared by the dataset metadata and OpenNeuro snapshot. - Requested citation: Sun, H.; Olszko, A. V.; Singh, N.; Jangraw, D. C. (2026), “ROAMM: A Benchmark Dataset for Multimodal Human Attention Decoding and EEG-to-Text Modeling During Naturalistic Reading,” ICML 2026, . For checksum verification, put the three hash/file pairs above in a local `SHA256SUMS` file and run `sha256sum -c SHA256SUMS`. The S3 `versionId` pins each byte sequence; avoid the equivalent unversioned bucket URLs for stable fixtures. Privacy: the dataset is public under CC0, but it remains pseudonymized human data. Its versioned [participants table](https://s3.amazonaws.com/openneuro.org/ds007629/participants.tsv?versionId=fhRtz_tSLrfDUt72m5OHdmF90kvXvWJY) publishes exact age, sex, handedness, ADHD diagnosis, and reading-disability status. That combination is sensitive and can increase linkage risk. Keep test fixtures access-controlled as appropriate and never describe the data as anonymous. Current-plugin note: unsupported. A future adapter must resolve the header's relative data/marker references safely and count the aggregate triplet size, not treat `.vhdr` as a standalone signal file. ## 4. NWB: DANDI mouse ALM session This is a real extracellular electrophysiology session from a house mouse in a delay-response task. It is unusually small for NWB and therefore well suited to header/tree preview tests without fetching a multi-gigabyte recording. - Immutable asset download endpoint: . - Asset metadata, including both content URLs and digest: . - Dandiset/version metadata: . - Path: `sub-anm369962/sub-anm369962_ses-20170316.nwb`. - Size: **609,568 B** (0.58 MiB). - Repository-published SHA-256: `c35c73d0e2b2f3a88520605036fb181eb53d207f554431612c922940b70d2717`. - License: CC BY 4.0; attribution is required. - Citation: Economo, M. N.; Svoboda, K. (2022), *Mouse anterior lateral motor cortex (ALM) in delay response task* (version 0.220126.1855), DANDI Archive, . Verification: ```bash curl -L --fail \ https://api.dandiarchive.org/api/assets/d97954f2-6d80-4dda-9ecf-89c7c10fed0b/download/ \ -o sub-anm369962_ses-20170316.nwb printf '%s %s\n' \ c35c73d0e2b2f3a88520605036fb181eb53d207f554431612c922940b70d2717 \ sub-anm369962_ses-20170316.nwb | sha256sum -c - ``` Privacy: the asset metadata identifies the subject as a house mouse (`NCBITaxon:10090`), not a person, so human participant privacy is not at issue. Normal attribution and research-integrity obligations still apply. Current-plugin note: unsupported. Keep the file for a future NWB/HDF5 metadata tree or Python-worker test; do not route it through the NIfTI parser. ## Fixture policy Do not silently vendor these data into the package. A test-data fetch script or CI cache should download on demand, verify SHA-256 before use, preserve the source citation/license alongside the cache, and use an explicit opt-in for human data. For permanent CI fixtures, prefer the animal NWB file and the single EDF file; the BrainVision binary is the largest download and is better suited to an integration-test job. ## Repository-local corpus `scripts/download-real-data.sh` downloads the four curated fixtures above and also keeps several supplementary, version-pinned parser fixtures locally: - the OpenNeuro `ds000005` run-01 BOLD image and events table, giving the current NIfTI Adapter a real `64 × 64 × 34 × 240` four-dimensional input; - the PhysioNet Sleep-EDF Expanded pair `SC4001E0-PSG.edf` and `SC4001EC-Hypnogram.edf` (ODC Attribution 1.0), for mixed-sampling-rate PSG signals and sleep-stage annotations; - EEGLAB's official `eeglab_data.set/.fdt` sample and `testbva` BrainVision regression triplet, pinned to EEGLAB commit `420769355b4e754628edfb43da2eb02c72fb8333`; - a second small DANDI `000006` session, `sub-anm369963_ses-20170228.nwb`, under the same CC BY 4.0 terms. The EEGLAB repository's BSD-2-Clause file explicitly describes the core source code and does not separately grant terms for every sample recording. These EEGLAB files are therefore downloaded only into the ignored local test cache; do not redistribute them without confirming the data owner's terms. All local files, including derived uncompressed NIfTI copies, are checked by `scripts/real-data.sha256`.