Zoidberg

• Clean and structured code for (almost) perfect reproduction. There are still some random factors in play, e.g., a random selection of candidate sentences, shuffle order. However, you should reproduce results with this code with almost equal accuracy reported on paper. To check on the exact and messy code we used in research, go to the "early_release" folder. For train on exact data we had in our research, rather than generate from this repo, you can download "parlai_training_data.zip" from the server.
• Possible bugs may exist. If you spot one, report via git issues.
• In this version, we dropped many features we had in original code and keep bear minimal for what we had on paper. In a possible future release, we will add more features and flexibilities for pipeline control and injection.