chatgpt-comparison-detection

Why we included this project

HC3, the Human ChatGPT Comparison Corpus at the heart of this repository, pairs human-written answers with ChatGPT responses across English and Chinese. It comes from the paper "How Close is ChatGPT to Human Experts?" and was one of the first datasets built specifically for measuring how far model output drifts from human writing, with examples drawn from sources like ELI5, Wikipedia, and medical and finance Q&A. The repo also ships a couple of working detectors, a RoBERTa-based classifier and a variant using linguistic features, that give you a reasonable baseline for flagging synthetic text on a single passage or a question-answer pair. This is research code rather than a production API, so expect to fine-tune the models and treat the corpus itself as the reference material for your own evaluation.