DistilBERT-base — AI-Generated Text Detector (HC3)
Fine-tuned distilbert-base-uncased for binary classification of human-written vs AI-generated text.
Trained as part of an MSc AI dissertation at the University of the West of Scotland (UWS), 2025/26. Supervisor: Dr Tahir Mahmood | Student: Abdul Hannaan Mohammed (B00409227)
Performance
| Condition | F1 | Accuracy | Recall |
|---|---|---|---|
| Clean HC3 (test) | 0.9922 | 0.9948 | 0.9995 |
| Pegasus paraphrase attack | — | — | ASR 5.2% |
| QuillBot-style attack | — | — | ASR 19.0% |
| ChatGPT rewrite attack | — | — | ASR 6.8% |
| M4 cross-dataset | 0.4316 | 0.6325 | — |
Training
- Base model: (Sanh et al., 2019) — 40% fewer params than BERT
- Dataset: HC3 (Hello-SimpleAI/HC3) — 55,156 training samples, 70/15/15 split
- Epochs: 4 | LR: 2e-5 | Batch: 16
- Hardware: Google Colab T4 (15 GB VRAM)