pting126/hw1-hc3-detector

HW1 HC3 Detector

MiniLM fine-tuned to classify English HC3 answers as human (0) or ChatGPT (1).

• Baseline test accuracy: 84.4901%
• Fine-tuned test accuracy: 99.6144%
• Best validation accuracy: 99.4856%
• Trained for 10 epochs; selected epoch 6 using validation accuracy.
• AdamW, learning rate 2e-05, batch size 32, seed 42.
• Answer text only; truncate inputs to 256 tokens.
• HC3 revision: 4d0ff18143b5a7e1b1e79beb540c04549d1e59d3.
• Question-level 80/10/10 split before flattening answers; paired answers stay together.
• Split sizes: {'train': 37334, 'validation': 4666, 'test': 4668}.

The baseline uses logistic regression on frozen MiniLM sentence embeddings. Full metrics and confusion matrices are in metrics.json (rows are true labels; columns are predicted labels, ordered human, ChatGPT).

For inference, load this repository with AutoTokenizer and AutoModelForSequenceClassification. Tokenize with truncation=True, max_length=256, and padding=True; use model.eval() and torch.no_grad().