Commit History
-
Drop the What is deliberately NOT here section
The scope of a checkpoint mirror is readable from the table: it lists weights and nothing else. Enumerating the training data that was skipped, with its sizes, explained a decision to nobody who needed it.8fcef57 -
Trim the README: drop the NOASSERTION aside and the gated-link section
The NOASSERTION clause explained a GitHub implementation detail that a reader of this mirror has no reason to care about; what matters is what the licence files say, which is quoted directly beneath. The paragraph about DeepNovo's link needing a Google sign-in is gone too. The table already marks that entry as gated, and the mirror's purpose does not need restating three times in one page.ca4026f -
Drop the Trained on column from the provenance table
The table's job is provenance and licence: which file, whose, from where, under what terms. What a checkpoint was trained on is a property of the model rather than of this copy, and it is already recorded per checkpoint in the catalog's own table and on each method's page, which this README links to.c2f3ce1 -
Point at the catalog's checkpoint table, not the page that no longer exists
/checkpoints was deleted before it was ever published: it duplicated the cross-method table that now lives on the catalog itself, where it is built from the database on every render instead of by hand in a second repository. This README was still advertising it. The replacement link is better anyway, because that table covers EVERY known checkpoint for these methods and not only the ones mirrored here.21b9de8 -
Both DeepNovo repos have a licence; NOASSERTION was my misreading
GitHub reports NOASSERTION when its detector cannot CLASSIFY a licence, not when a repository has none, and I repeated that as 'no licence can be relied on'. Both repos carry a LICENSE file granting academic and non-commercial use: DeepNovo 'publicly available for non-commercial uses' and DeepNovo-DIA 'provided free of charge for academic and non-commercial use', each with copyright reserved. Redistribution is still not expressly granted, so the table now says what the licences actually say and the mirror rests on being non-commercial and attributed rather than on a permission that was never there to begin with.fb95145 -
Credit pi-PrimeNovo's own repository, not a discussion thread
Both pi-PrimeNovo checkpoints are published by PHOENIXcenter/pi-PrimeNovo: the MassIVE model in its top-level README and the phosphorylation model in pi-PrimeNovo-PTM/README.md, both MIT. The ProteoBench discussion is where I first saw the phospho link, which is not the same as where it is published, and a provenance table should name the latter. That README also describes the model better than the discussion did: it is fine-tuned from model_massive.ckpt on 2020-Cell-LUAD and predicts one extra token, B, for Phosphorylation (+79.97).33ce54a -
Mirror DeepNovo's weights, retrieved from behind the sign-in wall
The original Drive link is gated: it answers "We can't access this content right now. Try signing in to your Google Account", while the Kaiko paper cites it as publicly available. The repository owner fetched it from a signed-in session, so this copy is now the only route to those weights for an anonymous reader, which is the clearest case in this mirror for why it exists. translate.ckpt-48600, its .index and the checkpoint meta file, plus knapsack.npy.zip: not weights, but DeepNovo cannot run without it, so a checkpoint on its own would be unusable. NOT mirrored from the same folder: uniprot_sprot.yeast.fasta, the yeast.low.coon_2013 test spectra, and the training logs and decode output. Only what is needed to run the model.995f976
-
Provenance and licence table for the mirror
c7359d7