Add MOSES-aq HDF5 training corpus (5 shards)
data/moses_aq/hdf5 -- the ShEPhERD-2 training corpus, 5 files, 5,725,141,659 B. Path follows the repo's existing data/<corpus>/<variant>/ convention and sits beside the 100-molecule data/moses_aq/scaffold-test fixture, which is a generation reference and NOT a training set.
Source: moses_aq_hdf5.zip of Zenodo record 22798054 (doi 10.5281/zenodo.22798054), verified byte-exact before unpacking -- 1,158,779,058 B, md5 1f7e591f49edff31170082c54790cfec, both matching the record. Each shard's size then matched the zip's own central directory, and MolCraftDiff zoo publish --stage-only re-confirmed every file against the manifest sha256 before staging.
Verified on read-back with h5py: every shard carries ['charges', 'charges_lengths', 'molblocks'] and the five index to 1,584,651 molecules -- consistent with MOSES's 1,584,663-molecule train split, 12 dropped, presumably conformer or xTB failures.
REGISTERED AS UNPACKED SHARDS, NOT AS THE ARCHIVE, deliberately: assets.py has no extraction step, and upstream's loader (hdf5_dataset.py:49) is sorted(self.hdf5_data_dir.glob('*.h5')) -- it wants a DIRECTORY of shards, so a registered .zip could not be consumed through ${asset:...} at all. The shards are stored uncompressed inside the zip's deflate stream, hence 1.08 GB compressed against 5.33 GB on disk. Each shard is standalone and the loader globs whatever subset is present, so a partial fetch still trains.
Licence: MIT. Zenodo record reports license mit-license, open access; the underlying MOSES corpus is MIT (molecularsets/moses).
Opened as a PR rather than pushed directly: the contributing token has read-only access to this repo.