This folder contains the following files:

• squad.py Creates the triggers using the optimization procedure described in the paper.
• universal_adversarial_sentences.txt contains the adversarial sentences used in the paper. You can copy and paste these into the variable adv_token_idxs in squad.py to test out different attacks.

We have pre-trained models available for use (see squad.py for how to use a model archive):

• bidaf as described in the paper using GloVe word embeddings. The model archive is available here https://s3-us-west-2.amazonaws.com/allennlp/models/bidaf-glove-2019.05.09.tar.gz