GlyCorpus: revised records and training roles

Published code and data are available at Zenodo DOI 10.5281/zenodo.23073476.

GlyCorpus contains 285,050 records with source identifiers, exact canonical WURCS strings, curation flags, sequence-index membership, pretraining partitions and benchmark-exclusion labels. schema.json defines the columns and count_flow.csv records the denominators.

Role	Records
Runtime pretraining training	221,827
Runtime pretraining development	55,440
Benchmark identity exclusions	3,833
Curation hold	3,841
Overlength exclusions	108
Absent from the frozen sequence index	1

Before length filtering, the training/development partitions contained 221,899 and 55,476 records. The 108 overlength exclusions comprise 72 training and 36 development records. Benchmark exclusions comprise 3,673 exact WURCS matches and 160 additional verified graph-identity matches.

Companion files contain MS acquisitions, corrected binding targets and dataset roles, and available benchmark and completion results. The MS collection contains 225,056 acquisitions from 3,035 parent glycans. Source labels distinguish the inputs used from the larger evidence inventory. The manifest and source records define the included assets; they should not be read as a complete archive of every raw trajectory or structure file.

The historical Supplementary File 1 is separate from this revised corpus. Licensing is documented for each asset.

Licensing

Author-generated records, features, coordinates, trajectories and annotations use CC-BY-4.0. Inherited inputs retain their source-specific terms and attribution; the CC-BY-4.0 license does not replace those terms. See TERMS.md and the source-specific data records.