radagon213/hw1-hc3-detector

HW1 — HC3 AI-generated text detector

sentence-transformers/all-MiniLM-L6-v2 fine-tuned end-to-end (AutoModelForSequenceClassification, 2 labels) on the English HC3 corpus (revision 4d0ff18143b5a7e1b1e79beb540c04549d1e59d3) to classify answers as 0 = human or 1 = ChatGPT.

Training setup

• Question-level 80/10/10 split, seed 42 (balanced: one human + one ChatGPT answer per question)
• AdamW, lr = 2e-05, batch size = 32, epochs = 5, max length = 256

Results (held-out HC3 test split)

Model	Test accuracy
Baseline: frozen all-MiniLM-L6-v2 embeddings + logistic regression	0.8449
Fine-tuned (this model)	0.9949

Limitations

HC3 is a historical benchmark (Dec 2022 ChatGPT). This model is not a reliable detector of text from current LLMs or of student work.