EasonListen/hw1-hc3-detector

HC3 Human–ChatGPT Answer Classifier

This model was fine-tuned from sentence-transformers/all-MiniLM-L6-v2 to classify English HC3 answers as human-written (label 0) or ChatGPT-generated (label 1).

Training used the homework’s provided question-disjoint 80/10/10 splits with seed 42, five epochs, batch size 32, learning rate 2e-5, and maximum sequence length 256 tokens.

Test results

• Frozen sentence-embedding + logistic-regression baseline accuracy: 84.49%
• Fine-tuned classifier accuracy: 99.61%
• Fine-tuned test errors: 18 out of 4,668 answers

The model was evaluated on historical HC3 data. These results do not establish reliable detection of writing from newer AI models or other domains.