DL20Questions is a repository for an implementation of a Natural Language Processing (NLP) system that plays the game 20 Questions. This project aims to demonstrate the capabilities of NLP in understanding and generating human-like questions in the context of the game.

• local-datasets: This folder contains the datasets we have created and used to train our models.
• fine-tuned-models: This folder contains our trained models.
• models-code: This folder contains the Python scripts for training and using our NLP models. These scripts are used to fine-tune and analyze our models.
• utils: This folder contains the script used to preprocess the data from the datasets mentioned above.
• "home" directory:
  • game-flow.py: This is the main script that uses the models and allows us to play the game with the artificial player.
  • DL20Questions_Report.pdf: Our final project report.
  • requirements.txt: The packages required for running our code.

To run the game, use the script game-flow.py located in this directory:

python game-flow.py

Before running the game, you need to train the GPT2 (Asker) model from the models-code directory:

cd models-code
python GPT2_ft2prompts.py

Optional (not required for running the game as the repository contains a checkpoint): For training the FLAN-T5 (Predictor) model, run the script flan-t5-LoRA.py from the models-code directory:

cd models-code
python flan-t5-LoRA.py

Eventually, all models will be located in the fine-tuned-models directory.

• GPT2_Inference.py:
  Inside the models-code directory there is a file named GPT2_Inference.py which is used to test the questions asker (GPT2) and provide perplexity analysis. In order to run it make sure to run GPT2_ft2prompts.py first, which generates the fine tuned GPT2 model.

cd models-code
python GPT2_Inference.py

Game teaser: