RITA ๐งฟ
Collection
A suite of autoregressive generative models for protein sequences, with up to 1.2Bparameters, trained on over 280 million protein sequences.
โข
4 items
โข
Updated
RITA is a family of autoregressive protein models, developed by a collaboration of Lighton, the OATML group at Oxford, and the Debbie Marks Lab at Harvard.
Model | #Params | d_model | layers | lm loss uniref-100 |
---|---|---|---|---|
Small | 85M | 768 | 12 | 2.31 |
Medium | 300M | 1024 | 24 | 2.01 |
Large | 680M | 1536 | 24 | 1.82 |
XLarge | 1.2B | 2048 | 24 | 1.70 |
For full results see our preprint: https://arxiv.org/abs/2205.05789
Instantiate a model like so:
from transformers import AutoModel, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("lightonai/RITA_xl, trust_remote_code=True")
tokenizer = AutoTokenizer.from_pretrained("lightonai/RITA_xl")
for generation we support pipelines:
from transformers import pipeline
rita_gen = pipeline('text-generation', model=model, tokenizer=tokenizer)
sequences = rita_gen("MAB", max_length=20, do_sample=True, top_k=950, repetition_penalty=1.2,
num_return_sequences=2, eos_token_id=2)
for seq in sequences:
print(f"seq: {seq['generated_text'].replace(' ', '')}")
@article{hesslow2022rita,
title={RITA: a Study on Scaling Up Generative Protein Sequence Models},
author={Hesslow, Daniel and Zanichelli, Niccol{\'o} and Notin, Pascal and Poli, Iacopo and Marks, Debora},
journal={arXiv preprint arXiv:2205.05789},
year={2022}
}