ESM-2 pretraining dataset

[jasperhyp]

Hi! I noticed in the paper that you used UniRef Sep 2021 release. But in their FTP (https://ftp.uniprot.org/pub/databases/uniprot/previous_releases/) there is no 2021-09 release. Can you clarify? Thanks!

[halilakin]

Hi. The dataset we used was created from 2021_04 release. Sorry for the confusion here.

[jasperhyp]

Thanks! Is the train/val split for this dataset uploaded? I only saw the split for the 2018 release used for ESM-1b. If I am not mistaken, the reported numbers of ESM-2 are on the val set, so I wanted to keep everything consistent!

[robert-verkuil]

(Adding some details re: the release date)
Link we referenced when writing "September 2021": Link
Just checked relnotes.txt, and it says "November 2021": Link

Either way it should be the 2021_04 release Halil's links above. 👍
(YYYY_NN seems to denote (year, release number), not (year, month).)

[jasperhyp]

Thank you! I know I'm asking for a lot but just one question more, is there possibly a fasta UniRef90 or UniRef50 that can be shared? The previous releases only contain several hundred GB+ XML files, which could take quite a while to parse...