Replace the four-source test MSA archive with one pre-merged and truncated A3M per protein. Add deterministic build and verification provenance plus per-file manifests; remove the obsolete folder index.
10 days ago
Checksums for the backfilled MSAs, index and excluded-protein list
20,423 entries, up from 14,157: the 6,264 extracted msas_test/ alignments plus the index and the excluded-protein list, following the existing convention of listing an archive's contents rather than the archive.
Cover msas_test/ and msas_test_index.parquet in the checksum sweep
- Adds an MSA section: the tar layout, the index, and the four-to-one merge in full (order, both caps, and the UniProt-vs-pooled deduplication), matching feature_provenance.msa in the model repo's config.json.
- Adds 'Matching the paper's numbers': 1,559 / 1,279 / 1,275 targets and which is which, since reporting one against another's target set shifts the headline by ~0.01.
- 'What is not here' no longer claims nothing else is missing: validation and train MSAs still are, and the test MSAs now ship.
The four test proteins whose sequences carry non-standard residues (UNK in the fold) that VN-EGNN cannot read, which the paper's comparison tables drop from every method. They are valid benchmark targets and site_policy.parquet scores them; the list is needed only to reproduce a table that includes VN-EGNN.
Four per-database alignments per test protein (UniProt, UniRef90, reduced BFD, MGnify), as shipped in the original upload. The paper promises the test MSAs and the release did not carry them; the README's merge recipe describes how the four become the single alignment ESMFold2 reads.
