HW1: HC3 AI-Generated Text Detector
Fine-tuned all-MiniLM-L6-v2 for binary classification of answers from HC3 (English).
Labels: 0 = human, 1 = ChatGPT.
Training
- Full fine-tuning, AdamW, lr 2e-5, batch size 32, 5 epochs, max length 256, seed 42
- Splits: 80/10/10 by question (pinned HC3 revision), train size 37,334
Results (HC3 test set, 4,668 answers)
| Model | Test accuracy |
|---|---|
| Baseline: frozen embeddings + logistic regression | 0.8449 |
| Fine-tuned | 0.9906 |
Limitations
Trained on a 2022 ChatGPT benchmark with known dataset artifacts. Not a reliable detector for current AI-generated or student text.
- Downloads last month
- 16
Model tree for jainatharva21/hw1-hc3-detector
Base model
nreimers/MiniLM-L6-H384-uncased Quantized
sentence-transformers/all-MiniLM-L6-v2