modernbert-ai-text-detector
Recommended general-purpose checkpoint: best cross-dataset transfer.
Fine-tuned ModernBERT-base classifier for detecting machine-generated English text (label 1 = machine). Built in the open research project robust-ai-text-detection, where every number regenerates from committed artifacts and the full research log, literature review, and negative results live.
Note. Strong cross-dataset transfer (M4GT), but weak on a 2025-generator probe in RAID's continuation format (TPR@1% 0.13) — no single checkpoint is robust on every axis; see the repository's research log and consider the companion raid-mix variant or a zero-shot ensemble for coverage.
Metrics
Thresholds must be calibrated on human text from your own distribution; the scores below use per-dataset calibration. AUROC alone is misleading for this task — use the low-FPR columns.
| eval | AUROC | TPR@5%FPR | TPR@1%FPR |
|---|---|---|---|
| MAGE test (near in-distribution) | 0.980 | 0.930 | 0.845 |
| HC3 (cross-dataset) | 0.994 | 0.970 | 0.941 |
| RAID eval grid, incl. attacks (OOD) | 0.875 | 0.512 | 0.310 |
| M4GT (cross-dataset) | 0.920 | 0.783 | 0.674 |
| Frontier Qwen3-2025 probe (cross-generator) | 0.881 | 0.297 | 0.126 |
Training
MAGE train (318K rows after artifact filtering), 1 epoch, lr 3e-5, 512 tokens.
Intended use and limitations
Research use. Output is probabilistic evidence, never proof of authorship: at 1% FPR, an institution processing 75,000 documents a year would wrongly flag ~750. Do not use as the basis of disciplinary action. Untested on non-native-writer false-positive rates; English only; document-level only; defeated by adaptive paraphrase attacks. Pair with Unicode/NFKC input normalization (see the repository) — it neutralizes homoglyph and zero-width attacks for free.
Citation
Please cite the datasets and methods this builds on (MAGE, RAID, ModernBERT — full BibTeX in the repository).