README.md

Workflow-guided Response Generation

Installation

git clone [email protected]:asappresearch/workflow-response.git
cd workflow-response
pyenv virtualenv workflow-response
pyenv activate workflow-response

Install the required packages:

pip install -r requirements.txt

Data Processing

Download and create datasets for training.

for dataset in abcd multi_woz; do
    bash scripts/dataproc/download_process_${dataset}.sh
done

Training

Finetune base model

bash bash/gpt2_train.sh

Train reward model

bash bash/bert_reward/train.sh

Train RL model

bash bash/quark_run.sh

Evaluation

Evaluate RL model:

python eval/interactive_quark_eval.py

Human evaluations:

python eval/process_human_eval.py