christofid/dapscibert

dapSciBERT

DapSciBERT is a BERT-like model trained based on the domain adaptive pretraining method (Gururangan et al.) for the patent domain. Allenai/scibert_scivocab_uncased is used as base for the training. The training dataset used consists of a corpus of 10,000,000 patent abstracts that have been filed between 1998-2020 in US and European patent offices as well as the World Intellectual Property Organization.