Pretokenized version of OMGprot50 in uint8 (ESM2 tokens).

Thanks Tatta Bio for putting together such an amazing base dataset!

Each shard (.bin) is 100,000,000 tokens

There are also added validation sets and test sets for evalution purposes, including ESM2 speed runs. OMG prot50 was clusterd at 50% identity, so random splits are nonredundant to the training set by default. Random splits of 10,000 make up the base components of the validation and test sets. To the test set, we also add all new Uniprot entries since OMG creation that have transcript level evidence after dedpulication.

Scripts

To download

Split creation

Tokenization