This repository contains the code for the paper Diverse and Effective Synthetic Data Generation for Adaptable Zero-Shot Dialogue State Tracking.
The domain-diverse D0T dataset can be found under data/dsg5k/train and is formatted as 3 CSV-of-JSON tables containing the dialogue turns (turn.csv), slot definitions (slot.csv), and slot-value pairs (slot_value.csv).
The main experiment script is dextrous/experiment.py (experiments presented in the paper launched this script as slurm jobs using launcher scripts under dextrous/launch).