This repository contains code to evaluate HuggingFace Seq2Seq models on the TOP family of datasets along with Constrained Adversarial Alignment.
Feel free to contact William Held with any questions at wheld3 [@] gatech edu.
For a description of methods, as well as reported results combining AMBER with pre-training alignment - see the [Paper]
pip install -r requirements.txt
Non-Adversarial: bash eval_t5.sh
Adverserial: bash eval_t5_adv.sh
Non-Adversarial: bash eval_t5_base.sh
Non-Adversarial: bash eval_t5_base_adv.sh
Non-Adversarial: bash train_t5.sh
Adverserial: bash train_t5_adv.sh
Non-Adversarial: bash train_t5_base.sh
Non-Adversarial: bash train_t5_base_adv.sh
alignment_mixin.py