Linda-Pro 1.1

Not a plain transformers model. Use the linda_pro package in this repo (windowing, calibration and verdict logic live there). Loading a models/* folder on its own gives weaker, uncalibrated scores.

Licensing in short: everything here (app, command-line tools, models, calibration) is free for personal non-commercial use (CC BY-NC 4.0). Commercial use needs one commercial licence that covers the app, the command-line tools and the models together: see LICENSING.md and COMMERCIAL_LICENSE.md. OEM, exclusive rights, custom calibration: lindapro.support@proton.me.

Version 1.0 (Linda-Essay v3, Linda-Multi v2, stylo7e) stays available in this repository's history (tag v1.0).

Runs on your own machine (GPU recommended, CPU works but the two transformer voters are slow). No cloud, no API, no data leaves the machine. Components: stylometry Stylo-D (LightGBM, CPU), Linda-Essay-D (DeBERTa-v3-large), Linda-Multi-D (mDeBERTa-v3-base). Text is cut into ~300-word windows (up to 12); each transformer scores every window, the text score is the mean of the top quarter of windows.

What is new in 1.1

The three voters were continued from 1.0 on data that now includes real outputs of commercial AI humanizers (two public benchmarks, CC BY 4.0) and open-model texts rewritten in humanizer styles. On HumanizerBench (CC BY 4.0, not used for training), June-September 2026, 1,675 humanized texts: 1.0 flagged 31% as ai, 1.1 flags 70% (sensitive mode, same 0.5%-false-positive threshold; with uncertain included 1.1 is higher still). Human texts were not flagged more often overall (TOEFL essays 8.8% -> 4.4%); school ELL essays 0.7% -> 1.1% and adult ESL (W&I) 1.1% -> 1.8% rose slightly, within sampling error. Details and caveats: EVIDENCE_PACK.md section 8.

Windows app (no Python needed)

Download Linda-Setup.exe from the GitHub releases and install it (no administrator rights needed). On first start the app downloads the files of this repository once (about 2.9 GB), then works offline, shows a colour map of the text and checks for model updates by itself. Free for personal non-commercial use. A commercial licence key ($10 person, $30 team of up to 10, $500 per year organization or university) covers the app, the command-line tools and these models together; details in LICENSING.md.

Speed

The whole ensemble on one text, measured on one machine (AMD Ryzen 7 7800X3D, 6 CPU threads; GPU: AMD Radeon RX 9070 XT through DirectML in the Windows app). Older laptops are slower. Scores on CPU and GPU agree to the third decimal.

text length	GPU (DirectML)	CPU, colours per ~300-word window	CPU, colours per sentence
300 words	0.14 s	1.1 s	3.8 s
1,000 words	0.34 s	3.5 s	8.1 s
3,000 words	0.98 s	11.9 s	27 s

Model loading at start takes about 5-10 s.

Quick start

git lfs install && git clone https://huggingface.co/Lindarixon/Linda-Pro && cd Linda-Pro
pip install -r requirements.txt
python -m linda_pro your_text.txt                # verdict + per-file summary
python -m linda_pro your_text.txt --windows      # per-window scores (mixed authorship)
python -m linda_pro your_text.txt --mode precise # fewer false accusations (schools, universities)

from linda_pro import LindaPro
det = LindaPro(mode="sensitive")          # or "precise"
r = det.detect([open("essay.txt", encoding="utf-8").read()])[0]
print(r["verdict"], r["ai_share"], r["windows"])

Local HTTP service: python -m linda_pro.server --port 8080 --token <secret> (DEPLOYMENT.md). Data flow, integrity check, dependencies: SECURITY.md. Verdict: ai / uncertain / human; ai_share = share of words in windows that look AI-generated (rough, resolution ~300 words).

Modes

• sensitive (default): ai if Linda-Essay OR the 3-voter ensemble exceeds its 0.5%-false-positive threshold; uncertain above the 5% threshold.
• precise: ai only if Linda-Essay is above its 1% threshold AND stylometry is above its 5% threshold. Fewer false accusations (TOEFL essays 0.0% in 1.1), catches less on essays by modern models and on humanizer output.

Measured quality (details and caveats: EVIDENCE_PACK.md)

Share of ai verdicts, sensitive / precise: HumanizerBench (all 1,675 humanized texts) 69.9% / 57.7%; AI Humanizer Benchmark October 2026 (363 humanized texts, held out) 81.8% / 68.0%; AI texts from three generators never seen in training (889) 93.0% / 79.4%; MAGE subset (1,400 AI texts, short, many old small models) 44.5% / 20.2% (was 50.5% / 27.6% in 1.0). False ai verdicts on held-out humans: news/reviews/blogs/books 0.2% / 0.1%, school ELL essays 1.1% / 0.2%, adult ESL (PELIC) 0.7% / 0.4%, (W&I) 1.8% / 0.0%, TOEFL exam essays 4.4% / 0.0%, long texts 0.0% / 0.0%, MAGE humans 1.1% / 0.1%. Public Chicago Booth benchmark (2025 humanizer version) at 1% false positives: plain AI 99.7%, after StealthGPT 82.9% (1.0: 85.4%; Pangram 98.1, GPTZero 44.3, Originality 29.1 — vendor numbers from the benchmark authors' repo).

Limits

English is the main language; Russian and Polish work at moderate quality (the network was trained on them, but thresholds are tuned for English: on our test sets the AI verdict catches about 63% of Russian and 51% of Polish AI texts at 0-1% false flags); treat those results as indicative. No comparison with other detectors was measured for these languages. Not better than Pangram. Humanizer tools change every month: the detector must be retrained on fresh outputs regularly. Weaker on texts of old small open models (MAGE) than 1.0. Essays by adult non-native writers (TOEFL) are still the main fairness risk in sensitive mode; use precise where a false accusation is costly. Never use as the sole basis for academic-misconduct or employment decisions. Data provenance: all three voters are Tier C (trained partly on outputs of commercial language models and AI humanizer tools); the stylometry voter is no longer strict-clean (the strict-clean stylometry remains available separately as Linda-Stylo-Clean). See DATA_BOM.md and LICENSE_COMMERCIAL_TEMPLATE.md. Requirements: Python 3.10+, torch, transformers, lightgbm, scikit-learn, numpy, pyyaml, sentencepiece, tiktoken (requirements.txt). Integrity: CHECKSUMS.sha256.

Citation

Technical report for version 1.1: Zenodo, DOI 10.5281/zenodo.23080472. The report for version 1.0 is 10.5281/zenodo.23072494.

Contact

Commercial licence, custom calibration, questions: lindapro.support@proton.me