Datasets:
Data provenance and terms
Raw MNIST and Fashion-MNIST archives are staged in data/, with exact download URLs, byte counts and SHA256 digests in data/archive_manifest.json. Archives were gzip-integrity checked. No model has evaluated the official test images. Archive retrieval is separate from final benchmark permission/locking.
| Dataset | Official source used | Terms evidence |
|---|---|---|
| MNIST | Google-hosted MNIST archive mirror | Keras dataset documentation identifies Yann LeCun and Corinna Cortes and states CC BY-SA 3.0. The original homepage is MNIST. |
| Fashion-MNIST | Zalando Research repository | The repository's MIT license identifies Copyright 2017 Zalando SE. |
| CIFAR-100 | Author-hosted dataset page and linked Python archive | The official page requests citation of Alex Krizhevsky, Learning Multiple Layers of Features from Tiny Images (2009). It does not specify a standalone dataset license; none is invented here. |
The bounded first CIFAR-100 download attempt exceeded 90 seconds and its partial archive was deleted. No CIFAR-100 archive/model evaluation is currently available from that attempt. The official Python archive MD5 advertised by the author is eb9058c3a382ffc7106e4002c42a8d85; it must be checked if the archive is retrieved later.
The baseline package and report should normally redistribute retrieval code and digests rather than embed dataset archives. Dataset terms are separate from any license of newly written research code; this note does not invent a project license.