hongchi/hw1-hc3-detector

HC3 human versus ChatGPT answer classifier

Binary labels: 0 = human; 1 = ChatGPT. Input is the answer text alone. Fine-tuned all-MiniLM-L6-v2 for five epochs on the English HC3 corpus. Questions were split 80/10/10 with seed 42 before answer pairs were flattened. The test split was held out from training and model selection.

Held-out test results

• Frozen-embedding logistic-regression baseline accuracy: 0.8449
• Fine-tuned classifier accuracy: 0.9919
• Fine-tuned macro F1: 0.9919

Limitations

This model distinguishes sources in a historical HC3 benchmark. It has not been validated on current models, domains, or student writing and should not be used to judge authorship of an individual text. Answers are truncated at 256 tokens.

Source: Guo et al. (2023), arXiv:2301.07597.