Repository navigation
A keyless, model-based check that is independent of the zh-style meter, so a rewrite can be judged by something it was not optimised against.
Sidecar
- New
fast-detectgptdetector (Bao et al., ICLR 2024, analytic criterion).Qwen/Qwen3-0.6B-Basesamples,Qwen/Qwen3-0.6Bscores, first 1024 tokens, lm_head applied in slices so both models fit beside the 4B model on a 12 GB card (about 3 GB more). uncertainat criterion ≥ 1.9, elseclean; fewer than 100 tokens →not_tested. Neverdetected: it is a heuristic, not a watermark.requires_keyis false.
UI
- Inspector row "Fast-DetectGPT (zero-shot)", flagged heuristic, so it never moves the watermark verdict.
- Rewrite panel: before/after Fast-DetectGPT score next to the zh-style meter when the sidecar has it.
Measured (46 human Chinese texts: 23 pre-2022 zh.wikipedia excerpts and 23 news wire articles from October 2026; 40 model-written zh-TW articles)
| Sampler → scorer | AUC |
|---|---|
| Qwen3-4B-Instruct-2507 → itself | 0.54 |
| Qwen3-0.6B → itself | 0.75 |
| Qwen3-0.6B-Base → itself | 0.85 |
| Qwen3-0.6B-Base → Qwen3-0.6B (shipped) | 0.955 |
| Qwen3-1.7B-Base → Qwen3-1.7B | 0.955 |
- At 1.9: 0/46 human false positives (max 1.85), 29/40 model texts flagged.
- Rewrites: 29/40 flagged before, 28/40 after deep, 26/40 after deep + recheck; median 2.18 → 2.29. LLM rewriting removes the phrasing zh-style looks for, but the text stays as predictable to a language model.
Tests: 1896 passed against upstream 5c18093.