prot_t5_xl_half_uniref50-enc EU AI Act Compliance Profile

Encoder only ProtT5-XL-UniRef50, half-precision model

An encoder-only, half-precision version of the ProtT5-XL-UniRef50 model. The original model and it's pretraining were introduced in this paper and first released in this repository. This model is trained on uppercase amino acids: it only works with capital letter amino acids.

Model description

ProtT5-XL-UniRef50 is based on the t5-3b model and was pretrained on a large corpus of protein sequences in a self-supervised fashion. This means it was pretrained on the raw protein sequences only, with no humans labelling them in any way (which is why it can use lots of publicly available data) with an automatic process to generate inputs and labels from those protein sequences.

One important difference between this T5 model and the original T5 version is the denoising objective. The original T5-3B model was pretrained using a span denoising objective, while this model was pretrained with a Bart-like MLM denoising objective. The masking probability is consistent with the original T5 training by randomly masking 15% of the amino acids in the input.

This model only contains the encoder portion of the original ProtT5-XL-UniRef50 model using half precision (float16). As such, this model can efficiently be used to create protein/ amino acid representations. When used for training downstream networks/ feature extraction, these embeddings produced the same performance (established empirically by comparing on several downstream tasks).

Intended uses & limitations

This version of the original ProtT5-XL-UniRef50 is mostly meant for conveniently creating amino-acid or protein embeddings with a low GPU-memory footprint without any measurable performance-decrease in our experiments. This model is fully usable on 8 GB of video RAM.

How to use

An extensive, interactive

Your obligations as a prot_t5_xl_half_uniref50-enc deployer

Track your progress

Sign up to track obligations, receive deadline reminders, and generate audit packages.

Tags

Library