DotCheck/valla-text-v14_3

Valla@14.3 is a text detector. It maps a UTF-8 string and an explicit lang code to (p \in [0,1]), an estimate of (P(\mathrm{AI})) for the writing, not for the truth of the claim.

Latin writing shares Rask and Comenius plus a named tiny per language. Chinese is Ma Jianzhong 2.2, independent of that trio. Ancestor Hub ids sit in the bases table and YAML base_model only.

• Latin group (en, es, pt, fr, it, de, nl): Rask + Comenius + named tiny (Sweet, Bello, Ribeiro, Littré, Ascoli, Grimm, Vries), then a 4-parameter stack (p=\sigma(w_0 z_r + w_1 z_c + w_2 \delta + b)) at cut 0.5.
• Chinese group (zh): Ma Jianzhong 2.2.

Unsupported lang fails closed (unsupported_language). There is no silent English fallback. Traditional-Chinese-dominant input is not scored (und).

Field	Value
Hub id	DotCheck/valla-text-v14_3
English wire	inhouse-text@14
Public token	Valla@14.3
English artifact	text_stack_head_v14_3_en.npz
Latin bases	Oxidane/tmr-ai-text-detector (MIT), intfloat/e5-small (MIT), prajjwal1/bert-tiny (MIT)
Rollback bases	fakespot-ai/roberta-base-ai-text-detection-v1 (Apache-2.0) — Valla rollback @12
Chinese base	hfl/chinese-macbert-base (Apache-2.0)
Heads	text_stack_head_v14_3_{en,es,pt,fr,it,de,nl}.npz + zh_v2_2
lang	en|zh|es|pt|fr|it|de|nl (pt uses the Portuguese 14.3 head; zh = Simplified Chinese)
Output	(p \in [0,1]) — estimated (P(\mathrm{AI}))
Serve	CPU FastAPI POST /v1/analyze-text with required lang

Model description

Text is cleaned (TC1) before hashing and scoring: wiki-style citations stripped, whitespace collapsed, #category tokens dropped. The same cleaner is used in Express and the extension.

Latin languages share one Rask forward and one Comenius forward per packed unit, then a language-specific tiny and stack. Chinese loads Ma Jianzhong only when text_stack_head_zh_v2_2.npz is present. Shared towers are not safe for concurrent forwards: the text lane serializes GEMMs.

In this repo: README.md, LICENSE, NOTICE, CITATION.cff, and the eight .npz heads listed above. Upstream encoder checkpoints are not redistributed here.

Architecture

text + lang
  → TC1 clean
  → if lang in {en, es, pt, fr, it, de, nl}:
        Rask + Comenius + named tiny
        (Sweet, Bello, Ribeiro, Littré, Ascoli, Grimm, Vries)
        → p = sigma(w0 zr + w1 zc + w2 delta + b)
  → if lang == zh:
        Ma Jianzhong
  → else: unsupported_language (no score)
  → p

Inference

Open weights: the live .npz heads in this repository (Apache-2.0), used with the frozen upstream bases named above. This is not a transformers AutoModel package.

Product scoring: Check or Pro API.

Minimum length and fair-use caps are product policy at Express, not properties of the heads.

Training data

English fit AI is commercial-clean self-generation (Qwen2.5-7B, Mistral-7B, and Grok-4.5). Holdout AI is a smaller disjoint generator (Qwen2.5-1.5B; encyclopedic-heavy). English humans are a warehouse sample: Gutenberg.org windows 40%, Wikipedia 35%, WikiText 25%, sliced to 520–1100 characters. Other Latin languages keep their prior wiki / WikiText / Gutenberg mixes. Chinese humans are Wikipedia, modern prose, and books (OpenCC t2s). Chinese fit AI is grok-4.6, Qwen2.5-7B, and Mistral-7B (encyclopedic / everyday / assistant_chat). Chinese exam AI is Yi-1.5-6B + SmolLM2 (encyclopedic; n=400 human / 480 AI). No NC datasets. No published holdout scraped from live ChatGPT or Claude.

	English	Other Latin	Chinese (zh-Hans)
Fit AI	Qwen2.5-7B + Mistral-7B + Grok-4.5	Qwen2.5-7B + Mistral-7B	grok-4.6 + Qwen2.5-7B + Mistral-7B
Holdout AI	Qwen2.5-1.5B	Qwen2.5-1.5B	Yi-1.5-6B + SmolLM2 (400/480 encyclopedic)
Human	Gut 40 / wiki 35 / WT 25	wiki / WikiText / Gutenberg	Wikipedia + modern + books (OpenCC t2s)

Evidence files: text_public_exam_v12.json (public EN), text_gates_v12.json (cousin jsonl, not public), text_gates_{lang}.json.

Evaluation

Absolute floors on every language head: human mean (P(\mathrm{AI}) \le 0.12), AI mean (\ge 0.85), balanced accuracy (\ge 0.90), all at threshold (0.5).

English (public claim)

Metric	Floor	Measured
mean (P(\mathrm{AI})) | human	(\le 0.12)	0.029
mean (P(\mathrm{AI})) | AI	(\ge 0.85)	0.984
balanced accuracy	(\ge 0.90)	0.979

Public English floats are RAID-lite overall (Wikipedia blocks; news is observed only). The frozen jsonl holdout is one-lab Qwen-1.5B and is not the public claim. Rank-first beat of @11 on that holdout still holds (BEAT_V11_HOLDOUT_OK). Canary/style v2 remain watches.

Other language heads (same floors)

Lang	Person	Wire	Human	AI	bal_acc
zh	Ma Jianzhong	inhouse-text-zh_v2	0.047	0.922	0.954
es	Bello	inhouse-text-es@14	0.029	0.993	0.982
fr	Littré	inhouse-text-fr@14	0.021	0.990	0.978
pt	Ribeiro	inhouse-text-pt@14	0.042	0.964	0.958
de	Grimm	inhouse-text-de@14	0.034	0.966	0.978
it	Ascoli	inhouse-text-it@14	0.016	0.943	0.972
nl	Vries	inhouse-text-nl@14	0.056	0.963	0.950

Portuguese lang=pt uses the Portuguese 14.3 head.

Intended use

• Reproduce the eight heads and the holdout tables.
• Research on synthetic-text detection in the eight supported languages.

Out of scope: languages outside {en, zh, es, pt, fr, it, de, nl}; Traditional Chinese; plagiarism, factual verification, or authorship; a product SLA from the Hub files alone.

Limitations

• Latin holdout AI is Qwen2.5-1.5B, disjoint from fit. Chinese exam AI is Yi-1.5-6B + SmolLM2 (encyclopedic). Live ChatGPT, Claude, and other closed generators are not a published holdout row.
• Short, heavily edited, or machine-translated text has higher variance on this stack.
• zh is Simplified Chinese (Ma Jianzhong). Other scripts and Traditional-Chinese-dominant text are out of domain: no score, no English fallback.

License

LICENSE — Apache License 2.0 for DotCheck heads in this repository. Upstream bases: NOTICE.

Citation

CITATION.cff · wire inhouse-text@14 / Valla@14.3 · https://dotcheck.ai/docs