transfer-learning-conv-ai

Why we included this project

This repo distills Hugging Face's ConvAI2 competition entry into roughly 250 lines of commented training code, making it a handy study tool for anyone learning to fine-tune a dialog model from a pretrained GPT or GPT-2. The scripts are readable end to end, cover distributed and FP16 training, and come with an interact script that loads the released checkpoint, so you can talk to the model right away. Training your own version takes about an hour on an 8 GPU V100 cluster, which is practical for a small team iterating on ideas. It is a useful baseline for understanding how transfer learning applies to conversation, though it is research code rather than a maintained production chatbot.