spikeforge/data/dataset_provenance.py records every registry dataset's upstream licence. Eleven of thirteen are now checked against the publisher's own page. Two are not, both because the source refuses to serve:
| Dataset |
Blocker |
mnist |
yann.lecun.com/exdb/mnist refused connection on every attempt (2026-09-14, 2026-09-15) |
cifar10_dvs |
figshare item 4724671 returns HTTP 403; its download endpoints return HTTP 202 with no body |
Both carry the unverified-candidate marker, and the entry's source field records UNCHECKED with the reason. The licences in circulation for both (CC-BY-SA-3.0 and CC-BY-4.0 respectively) are secondary sources, so neither is asserted.
MNIST is the one that matters. Four of the six shipped reference checkpoints train on it, so four catalog entries currently disclose an unverified data licence. Confirming it upgrades a third of the catalog.
To resolve: read the licence from the publisher's own page, then
# edit spikeforge/data/dataset_provenance.py, then:
python scripts/train_reference_models.py --sync-provenance
which rewrites dataset_license/dataset_attribution on every affected entry without retraining anything. Note that usps, emnist_digits, emnist_letters and cifar10 also carry the marker but are not part of this issue — those were checked and their publishers simply declare no licence, which the marker represents correctly.
spikeforge/data/dataset_provenance.pyrecords every registry dataset's upstream licence. Eleven of thirteen are now checked against the publisher's own page. Two are not, both because the source refuses to serve:mnistyann.lecun.com/exdb/mnistrefused connection on every attempt (2026-09-14, 2026-09-15)cifar10_dvsHTTP 403; its download endpoints returnHTTP 202with no bodyBoth carry the
unverified-candidatemarker, and the entry'ssourcefield recordsUNCHECKEDwith the reason. The licences in circulation for both (CC-BY-SA-3.0 and CC-BY-4.0 respectively) are secondary sources, so neither is asserted.MNIST is the one that matters. Four of the six shipped reference checkpoints train on it, so four catalog entries currently disclose an unverified data licence. Confirming it upgrades a third of the catalog.
To resolve: read the licence from the publisher's own page, then
# edit spikeforge/data/dataset_provenance.py, then: python scripts/train_reference_models.py --sync-provenancewhich rewrites
dataset_license/dataset_attributionon every affected entry without retraining anything. Note thatusps,emnist_digits,emnist_lettersandcifar10also carry the marker but are not part of this issue — those were checked and their publishers simply declare no licence, which the marker represents correctly.