DotCheck/valla-text-v9

Apache-2.0 text AI-likeness heads for DotCheck (English, Chinese, and six Latin language heads). This repo includes the live .npz heads, model card, license, and notices.

Field	Value
Hub id	DotCheck/valla-text-v9
Wire id (EN)	inhouse-text@9
Label	Valla
EN artifact	text_stack_head_v9.npz
Latin feature bases	Oxidane/tmr-ai-text-detector (MIT), fakespot-ai/roberta-base-ai-text-detection-v1 (Apache-2.0)
Chinese feature base	hfl/chinese-macbert-base (Apache-2.0; cls_emb)
Heads	EN + zh_v1 / es_v2 / pt_BR_v1 / fr_v2 / it_v1 / de_v1 / nl_v1 (separate .npz)
Output	p ∈ [0,1] — P(AI-like)
Serve	CPU FastAPI POST /v1/analyze-text + required lang
lang	en|zh|es|pt|fr|it|de|nl (pt → pt_BR head; zh = Simplified Chinese)

Model description

Latin group: Frozen TMR + Fakespot encoders → shared feature vector → language-specific logistic head.
Chinese group: Frozen MacBERT cls_emb → zh_v1 logistic head (train=serve contract).

Serve loads the Latin base pair for all Latin heads; the Chinese tower loads when text_stack_head_zh_v1.npz is present. Unsupported lang → fail closed (unsupported_language; no silent EN fallback). Traditional-Chinese-dominant product input is not scored (und).

Text hygiene (TC1) applied before hash/score: strip wiki-style cites, collapse whitespace, drop #category tokens (mirrored in Express / extension).

Files in this repo: README.md, LICENSE, NOTICE, CITATION.cff, and the .npz head file(s) listed above.

Architecture

raw text
  → text_clean (TC1)
  → if lang in {en,es,pt,fr,it,de,nl}:
        TMR + Fakespot frozen forward (shared Latin group)
  → if lang == zh:
        MacBERT cls_emb frozen forward (Chinese group)
  → lang-selected logistic head (npz)
  → p_AI

Inference

Open weights: the live .npz head(s) in this repo (Apache-2.0), for use with the frozen upstream backbone named above. This is not a transformers AutoModel.from_pretrained("DotCheck/…") package.

Product scoring: Check or Pro API (below). Leviathan (shared memory and related product path) is not in these files.

curl -sS -X POST "https://dotcheck-server-c221c1f32c68.herokuapp.com/analyze-text" \
  -H "Authorization: Bearer dc_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"text\":\"...\",\"lang\":\"en\"}"

UI: https://dotcheck.ai/check · API: https://dotcheck.ai/api · docs: https://dotcheck.ai/docs
Min length / FUP enforced at Express (product policy).

Training data

	EN / Latin	Chinese (zh-Hans)
Fit AI	Qwen2.5-7B + Mistral-7B (multi-temp)	Qwen2.5-7B + Mistral; wiki-style prompts
Holdout AI	Qwen2.5-1.5B	Qwen2.5-1.5B (~200/200)
Human	hard-neg wiki / WikiText / Gutenberg mix	Wikipedia lead prose (OpenCC t2s)
License	commercial-clean only (no NC banned sets)	same

No published holdout from live ChatGPT/Claude scrapes. Evidence: text_gates_v9.json, text_gates_{lang}.json.

Evaluation

English (public claim SSOT)

Metric	Target	Measured
mean P(AI) | human	≤ 0.12	0.045
mean P(AI) | AI	≥ 0.85	0.909
bal_acc @ thr	≥ 0.90	0.980

Also: beat prior @6 holdout; wiki-monitor mean ~<0.001; OOD / RAID-lite protocol OK (TEXT_GATES_OK).

Language heads (holdout; same absolute floors)

Lang	Wire	Human	AI	bal_acc
zh	inhouse-text-zh_v1	0.033	0.973	0.972
es	inhouse-text-es_v2	0.088	0.953	0.972
fr	inhouse-text-fr_v2	0.082	0.976	0.975
pt	inhouse-text-pt_BR_v1	0.103	0.945	0.928
de	inhouse-text-de_v1	0.064	0.968	0.958
it	inhouse-text-it_v1	0.069	0.986	0.972
nl	inhouse-text-nl_v1	0.101	0.978	0.933

Intended use

• Supported-language AI-likeness scoring in DotCheck inference.
• Citation of the tables above.

Out of scope

• Product scoring SLA / Leviathan / FUP via Hub download
• Languages outside {en,zh,es,pt,fr,it,de,nl}
• Traditional Chinese as a scored variant (fail closed)
• Plagiarism / authorship adjudication

Limitations

• Short, MT-heavy, or heavily edited text: higher variance.
• Eight languages only (Simplified Chinese for zh).
• Output is not generator attribution.

License

LICENSE — Apache License 2.0 for DotCheck heads in this repo. Upstream backbones: see NOTICE.

Citation

CITATION.cff · wire inhouse-text@9 / Valla@9 · https://dotcheck.ai/docs