isletyang/hw1-hc3-detector

HC3 answer classifier

This binary classifier fine-tunes sentence-transformers/all-MiniLM-L6-v2 on the English HC3 answer texts. Label 0 means human answer; label 1 means ChatGPT answer.

The data was split by normalized question into training, validation, and test sets with seed 42. One nonempty answer from each class was kept per question, yielding 37,334 training answers, 4,666 validation answers, and 4,668 test answers.

Held-out test results

• Frozen-embedding logistic-regression baseline accuracy: 0.8443
• Fine-tuned classifier accuracy: 0.9949 (24 errors out of 4,668 answers)

Training used five epochs, batch size 32, maximum sequence length 256, and AdamW at learning rate 2e-5. This is a result on the historical HC3 split, not a reliable detector for current student writing or other model families.

from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo = "isletyang/hw1-hc3-detector"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo)