Open-weights AI text detector ParaTrace v8 hits 99.2% TPR at 1% FPR, 5th on RAID leaderboard
AltruisticCouple3491 · reddit · 2026-11-13
Researchers open-sourced ParaTrace v8 (RAID-tuned), ranking 5th on the RAID leaderboard on TPR at 1% FPR.
Key details:
- Ettin-1B encoder (ModernBERT architecture), 512-token windows
- Fine-tuned on RAID's training split, followed by a hard-example round
- RAID hidden test set: TPR 99.20% at 1% FPR, 99.76% at 5% FPR (99.43% at 1% without attacks)
- Weakest slices: Cohere generations (96.5%), Reddit (97.9%), paraphrase attack (98.0%)
- detector.py reproduces the leaderboard submission to within 1e-6
Caveat from the model card: tuned specifically for RAID — on ordinary writing it misses much AI text and flags some human text, so it's for research, not for judging people. Released under CC BY-NC-SA 4.0 on Hugging Face.
More from Research
- MBZUAI's IFM open-sources PAN 2, a 29B world model that simulates action outcomes — LinusEkenstam · 2026-11-13
- 2008 study: people navigate first (56-68% of retrievals), search is only 4-15% — dSebastien · 2026-11-13
- Cyc: the 40-year quest to hand-code common sense that AI history forgot — ArtificialOther · 2026-11-13
- Open Vision Banana: fully open recipe beats Vision Banana with 73.9 mIoU on Cityscapes — FunAILab · 2026-11-13
- JustRUA: full episode history alone substantially boosts GPT-6 robot-agent scores — Zefan Wang · 2026-11-13
- Survey: why self-improvement mechanisms for multimodal agents transfer unevenly — Jingjing Jiang · 2026-11-13