Lindarixon/Linda-Pro

Linda Assay — local AI-text detector (English, Polish, Russian)

Formerly named Linda-Pro.

Free for personal non-commercial use, paid for commercial use - one licence covers the app, the command-line tools and the models (LICENSING.md, COMMERCIAL_LICENSE.md, contact lindapro.support@proton.me). Runs on your machine: no cloud, no API. Weights live on Hugging Face: Lindarixon/Linda-Pro (ensemble, GPU recommended) and Lindarixon/Linda-Stylo-Clean (CPU only, strictly clean data lineage).

Windows app (no Python needed)

Download Linda-Setup.exe from the releases page, install it and start it: the app downloads the models once (about 2.9 GB), then works offline, highlights the AI-looking passages and updates its models by itself. Free for personal non-commercial use; a commercial licence key ($10 person, $30 team of 10, $500 per year organization) covers the app, the command line and the models together (how licensing works). Source of the app: desktop/.

Two versions: Linda Assay and Linda Loupe

The same app, two sets of models. Install either one or both; the installer asks, and Settings let you change it later (the app also works with only one of them).

• Linda Loupe: a small, fast model for any computer, including old laptops without a graphics card. Slightly less accurate.
• Linda Assay: the full ensemble, the most accurate; best with a modern graphics card. Not sure? Take Lite: it runs everywhere, and Linda Assay can be added later.

	Linda Assay	Linda Loupe
Texts rewritten by 16 "humanizers" (HumanizerBench), found	81.0%	69.4%
Newer humanizer texts (Oct 2026), found	84.0%	75.8%
Texts of three unseen AI generators, found	95.5%	83.0%
Polish AI texts found (false alarms)	79.4% (1.4%)	72.2% (1.5%)
Russian AI texts found (false alarms)	80.3% (0.6%)	77.3% (1.2%)
False alarms, non-native exam essays (TOEFL)	5.5%	2.2%
False alarms, human texts of a mixed set (MAGE)	2.0%	0.4%
Check of a 3,800-word text	1-3 s with a modern graphics card; minutes on a processor alone	about 1-5 s on one ordinary processor (measured 1.2 s on a Ryzen 7, 4 threads)
Download	about 2.9 GB	about 0.3 GB
Graphics card	recommended	not needed

Recommended computer:

	Linda Loupe	Linda Assay
System	Windows 10 / 11, 64-bit	Windows 10 / 11, 64-bit
Processor	any 4-core processor from about 2017	4 cores or more
Memory (RAM)	4 GB (8 GB is comfortable)	8 GB (16 GB recommended)
Graphics card	not needed	DirectX 12 card with 4 GB of video memory or more; works without it, but slowly
Free disk space	2 GB	6 GB (an SSD is better)

Our own measurements on public test sets (English, Polish, Russian), not an independent audit; one calibration rule for both versions. The full Linda Assay needed 226 s for a 3,800-word text on an office Ryzen 5 PRO 2400GE; Lite is expected to take a few seconds there (not measured on that processor yet). Lite weights have their own repository: Lindarixon/Linda-Pro-Lite. From Python: python download_models.py --set lite, then python -m linda_pro file.txt --lite.

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 differ slightly (about 0.1 on a scale where the AI line is near 0); verdicts matched on our checks.

text length	GPU (DirectML)	CPU, quick look	CPU, careful look
300 words	0.08 s	0.8 s	2.1 s
1,000 words	0.25 s	2.7 s	5.7 s
3,000 words	0.76 s	8.4 s	20 s

Model loading at start takes about 5-10 s; the two networks need about 1.6 GB of GPU memory.

Test it on YOUR texts (nothing leaves your machine)

The weights are public on Hugging Face: Lindarixon/Linda-Pro (no login needed).

pip install -r requirements.txt
python download_models.py                              # downloads the 1.3.1 weights (~2.9 GB) from Hugging Face into ./models
python -m linda_pro my_text.txt --windows               # verdict + per-window scores (mixed authorship); language is detected automatically
python evaluate.py --human my_human_texts/ --ai my_ai_texts/ --mode precise   # accuracy on your own labelled samples

evaluate.py takes two folders of .txt files (or a JSONL with text and label fields), prints AUROC, true/false positive rates for the ai verdict and for ai or uncertain, and writes evaluation_report.csv. Use at least 100 texts per class, 150+ words each; texts by non-native writers are a good stress test. From Python: from linda_pro import load_detector; load_detector("sensitive").detect([text]).

No GPU? Google Colab

Open colab/Linda_Pro_demo.ipynb in Colab (free T4 is enough), run all cells, paste or upload your texts.

What it is

Version 1.3.1 (weights; Windows app 2.0.3.x). Three voters: stylometry (CPU) and two transformers, Linda-Essay-M (DeBERTa-v3-large) and Linda-Multi-M (mDeBERTa-v3-base), each a weight-interpolated "merge" of three fine-tuned checkpoints; one decision rule per language (English, Polish, Russian). The language is detected automatically. Version 1.3.1 was retrained without any text generated by Llama 3.x models; earlier versions (1.1, 1.3.0) used some, see DATA_BOM.md. Text is cut into ~300-word windows; verdict ai / uncertain / human, ai_share (rough share of AI-looking text) and per-window scores. Modes: sensitive (default) and precise (fewer false accusations for schools/universities). Results, limits and data provenance: EVIDENCE_PACK.md, DATA_BOM.md, MODEL_CARD.md.

Short version (sensitive mode, our own evaluation, not an independent audit):

• English: on HumanizerBench (not used for training) 81.0% of 1,675 humanized texts are flagged ai (1.1: 69.9%; 1.0: 31.4%); on the held-out October 2026 cycle of the AI Humanizer Benchmark 84.0%; on three unseen modern generators 95.5%.
• Public Chicago Booth benchmark: plain AI 99.8% (Pangram 99.9, GPTZero 98.6, Originality 94.2), after StealthGPT 88.0% (Pangram 98.1, GPTZero 44.3, Originality 29.1). We are not better than Pangram on humanized text.
• Russian and Polish (our own small test sets, 350 AI + 350 human texts each): the ai verdict catches 80.3% of Russian AI texts at 0.6% false flags and 79.4% of Polish at 1.4%; texts rewritten by another model to look human are caught 59% (Russian) and 56% (Polish) of the time. Sets are small (about +-5 points). No comparison with other detectors was measured for these languages.
• False flags on human texts: 0.2-0.4% on general, school and adult-learner English sets, but 5.5% on TOEFL essays by non-native writers and 2.0% on the MAGE human subset (1.3.0 had 0.9%). Precise mode lowers false flags (TOEFL 1.1% in 1.3.0; not re-measured for 1.3.1). Never use a verdict as the sole basis for decisions about people.

Deployment (hardware, offline install, local HTTP service python -m linda_pro.server): DEPLOYMENT.md. Data flow, integrity check, dependencies: SECURITY.md.

Commercial licence, custom calibration on your texts, integration

lindapro.support@proton.me — describe the use case, volume and languages.

Citation

If you use Linda Assay or Linda Loupe, please cite the technical report for version 1.3.2 (Zenodo, DOI 10.5281/zenodo.23261670). Earlier reports: version 1.3.1, 10.5281/zenodo.23242014; version 1.1, 10.5281/zenodo.23080472; version 1.0, 10.5281/zenodo.23072494.

@techreport{manzyuk2026lindapro132,
  title     = {Linda Assay 1.3.2 and Linda Loupe: two local detectors of AI-generated text in English, Polish and Russian},
  author    = {Manzyuk, Vladyslav},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23261670}
}