Exploiting fine tuning on GPT2 transformer for conversational AI

This thesis focuses on the development of a Conversational AI Chatbot that aims to retain conversation history and provide consistent responses. Two approaches are used: fine-tuning the DialoGPT model and multi-task learning with the basic GPT2 model. The conversational dataset preparation phase involves collecting relevant data from sources like chatbots and conversation logs. Two datasets, Persona-Chat and Cornell Movie Dialogs Corpus, are used. Training and testing involve using the prepared dataset to train the chatbots and evaluate their performance using the perplexity metric. The thesis draws inspiration from previous works, such as TransferTransfo, 'gpt2-chatbot' project, and 'How to Build a State-of-the-Art Conversational AI with Transfer Learning.' The goal is to develop an advanced Chatbot using Python, with the code available on GitHub and executable on Google Colab. The transformers, including GPT2 and DialoGPT, are accessible through the Hugging Face platform.