This model is based on RoBERTa and was pre-trained on 1.7 million sentences.

The training corpus was papers taken from Semantic Scholar's CORD-19 historical releases. Corpus size is 13k papers, ~60M tokens. I used the full-text "body_text" of the papers in training (details below).


from transformers import pipeline
from transformers import RobertaTokenizerFast, RobertaForMaskedLM

tokenizer = RobertaTokenizerFast.from_pretrained("amoux/roberta-cord19-1M7k")
model = RobertaForMaskedLM.from_pretrained("amoux/roberta-cord19-1M7k")

fillmask = pipeline("fill-mask", model=model, tokenizer=tokenizer)

text = "Lung infiltrates cause significant morbidity and mortality in immunocompromised patients."
masked_text = text.replace("patients", tokenizer.mask_token)
predictions = fillmask(masked_text, top_k=3)
  • Predicted tokens
[{'sequence': '<s>Lung infiltrates cause significant morbidity and mortality in immunocompromised patients.</s>',
  'score': 0.6273621320724487,
  'token': 660,
  'token_str': 'Ġpatients'},
 {'sequence': '<s>Lung infiltrates cause significant morbidity and mortality in immunocompromised individuals.</s>',
  'score': 0.19800445437431335,
  'token': 1868,
  'token_str': 'Ġindividuals'},
 {'sequence': '<s>Lung infiltrates cause significant morbidity and mortality in immunocompromised animals.</s>',
  'score': 0.022069649770855904,
  'token': 1471,
  'token_str': 'Ġanimals'}]


  • About
    • name: CORD-19: The Covid-19 Open Research Dataset
    • date: 2020-03-18
    • md5 | sha1: a36fe181 | 8fbea927
    • text-key: body_text
    • subsets (total: 13,202):
      • biorxiv_medrxiv: 803
      • comm_use_subset: 9000
      • pmc_custom_license: 1426
      • noncomm_use_subset: 1973
  • Splits (ratio: 0.9)
    • sentences used for training: 1,687,124
    • sentences used for evaluation: 187,459
  • Total training steps: 210,890
  • Total evaluation steps: 23,433


  • Data
    • block_size: 256
  • Training
    • per_device_train_batch_size: 8
    • per_device_eval_batch_size: 8
    • gradient_accumulation_steps: 2
    • learning_rate: 5e-5
    • num_train_epochs: 2
    • fp16: True
    • fp16_opt_level: '01'
    • seed: 42
  • Output
    • global_step: 210890
    • training_loss: 3.5964575726682155


  • Perplexity: 17.469366079957922


Allen Institute CORD-19 Historical Releases

    title={CORD-19: The Covid-19 Open Research Dataset},
    author={Lucy Lu Wang and Kyle Lo and Yoganand Chandrasekhar and Russell Reas and Jiangjiang Yang and Darrin Eide and K. Funk and Rodney Michael Kinney and Ziyang Liu and W. Merrill and P. Mooney and D. Murdick and Devvret Rishi and Jerry Sheehan and Zhihong Shen and B. Stilson and A. Wade and K. Wang and Christopher Wilhelm and Boya Xie and D. Raymond and Daniel S. Weld and Oren Etzioni and Sebastian Kohlmeier},
Downloads last month
Hosted inference API
Mask token: <mask>
This model can be loaded on the Inference API on-demand.