NLP Lab Course 2021: VDA Project

This is the repository for the NLP Lab Course 2021: Virtual Dietary Advisor.

1 Organization of the Repository

The repository structure is as follows:

• pre-training-gpt2/: It contains code to train GPT-2 on empathetic dialogue dataset in a supervised setting.
• annotate_data/: It contains code to annotate sampled responses from supervised GPT-2.
• train_reward_policy/: It contains code to train reward model and policy model in an RL setting.
• evaluate_models/: It contains code to evaluate our trained models with respect to different evaluation metrics.
• slides/: It contains our milestones and final presentation slides.
• report/: It contains final report of our project.

2 Instructions

• This code has only been tested on Google Colab.

• All our model checkpoints, annotated data and sampled responses are in google cloud storage. (gs://nlp-lab/)

• Read README.md of respective folder before running the code.

3 Running the code

• Clone the repository git clone https://gitlab.lrz.de/social-computing-research-group/nlp-lab-course/nlp-lab-course-ss2021/vda/nlp-2021-vda.git.

• Upload the entire folder nlp-2021-vda to google drive

• Open notebook of respective directory (XX/notebooks/) based on what you want to run.

4 Warning

• Our google cloud storage expires on 25th July, 2021, therefore, make sure to download models and datasets locally before that.