GPT2-chitchat

Why we included this project

This is a compact GPT-2 based Chinese chitchat bot that keeps the whole workflow in one repository. You get the data preprocessing script, the training code, and an interactive command-line script for chatting with the finished model, so you can move from a raw conversation corpus to something that talks back without assembling pieces yourself. The implementation sits on HuggingFace transformers and uses temperature, top-k, and nucleus sampling during generation to keep replies from degenerating into repetition. Code comments are in Chinese and quite detailed, which makes the project a useful study aid for learning how multi-turn dialogue fine-tuning fits together; the author also publishes the chitchat datasets used for training. Microsoft's DialoGPT project cites this work, and it's a reasonable prototyping base, though with dependencies pinned to an older stack (PyTorch 1.7, transformers 4.2) it is not something to drop straight into production.