Instructions to use empirischtech/paratrace-v10 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use empirischtech/paratrace-v10 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="empirischtech/paratrace-v10")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForSequenceClassification tokenizer = AutoTokenizer.from_pretrained("empirischtech/paratrace-v10") model = AutoModelForSequenceClassification.from_pretrained("empirischtech/paratrace-v10", device_map="auto") - Notebooks
- Google Colab
- Kaggle
You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
ParaTrace v10 is released for non-commercial research and evaluation under CC BY-NC-SA 4.0. Its verdicts are evidence, not proof: it flags some human writing and misses some AI text, so it must not be the only basis for a decision about a person.
Log in or Sign Up to review the conditions and access this model content.
ParaTrace v10
The AI-text detector behind paratrace.net since 14 November 2026. It is a 4B-parameter decoder that scores English text for how much of it is AI-written. For each text it returns a score, a verdict, a confidence level and a category (human, AI-assisted or AI-generated) and, on request, which parts of a long text look AI-written.
It replaces ParaTrace v7. On the same tests it catches far more humanized text (StealthGPT-humanized: 92.0% against 21.4% at 1% FPR) and flags less human writing (fresh human web text: 0.2% against 1.9%), for about four times the computation.
| Base model | Qwen/Qwen3-4B-Base (Qwen3 architecture, 4.0B parameters) |
| Input | English text, scored in 512-token windows |
| Output | the expected share of AI-written text (0 to 1); a verdict at the strict threshold (1% of human texts flagged) or the lenient one (5%); confidence; a category: human, AI-assisted or AI-generated; optionally labelled segments |
| Weights | bfloat16, the served model's scoring weights unchanged (the service runs them in float16 on a Quadro RTX 8000) |
| License | CC BY-NC-SA 4.0, non-commercial (why); access after accepting the terms |
Performance at a glance
RESULTS.md has every table and the method.
| Test | Result |
|---|---|
| Chat-assistant answers: 25,601 from 90 models | 99.8% caught |
| Answers from late-2025 models not in its training data: 1,116 | 99.4% caught |
| AI text under 11 evasion attacks: 16,500 texts | 95.9% caught |
| AI-written poems, a genre held out of training: 5,000 | 90.4% caught |
| Held-out human writing: 22,681 texts from 43 sources | 1.04% flagged |
| Fresh human web text: 66,586 texts | 0.2% flagged |
| UChicago human controls: 7,968 texts | 0.10% flagged |
| StealthGPT-humanized AI text (UChicago): 15,932 texts | 92.0% caught at 1% FPR |
| AI-improved student essays (GEDE): 285 | 98.95% caught at 1% FPR |
| LLM paraphrases of human text: 3,000 | 98.4% labelled AI-assisted |
| RAID, local estimate on held-out documents | TPR 92.4% at 1% FPR; AUROC 99.10 |
"Flagged" and "caught" are at the strict threshold, the one the service uses by default. The first five rows are ParaTrace's benchmark report of the served model (float16, through the production API); the others are development evaluations of the same model (bfloat16).
Known weaknesses
- Persuasive essays. Human persuasive essays often look AI to it: 20.1% are flagged (35 of 174). They were kept out of training.
- Humanized text it was not trained on. Of AI texts rewritten by a humanizer model held out of training, 45.6% are caught. Short humanized texts are harder still: 36.4% of StealthGPT-humanized texts under 50 words are caught at 1% FPR.
- Edited human writing. Human text paraphrased by an AI is flagged 52.2% of the time, and 70.6% of the flagged texts are called AI-generated rather than AI-assisted. Human text run through a synonym swapper is flagged 37.6% of the time.
- Some genres: recipes (5.8% flagged), book excerpts (3.8%), speeches and debates (3.7%).
- Short texts. Under 50 words, 4.1% of human texts are flagged, and the result carries a warning.
- New domains and base models are caught less often: poetry 90.4%. On RAID, text from Cohere's base model is caught 70.3% of the time, and paraphrased text 83.3%.
- Compute. About four times v7's computation per word, and 8.3 GB of GPU memory: 161 ms for a 300-word text on a Quadro RTX 8000.
- English only.
How to use
Accept the terms on this page; access is granted at once.
Log in and download the scoring code:
pip install torch "transformers>=4.57" huggingface_hub hf auth login hf download empirischtech/paratrace-v10 detector.py --local-dir .Score texts:
from detector import Detector detector = Detector.from_hub() # downloads the model (8 GB); or Detector("local/folder") detector.predict(["Text to check ..."]) # strict threshold # [{'probability': 0.9999, 'is_ai': True, 'operating_point': 'strict', 'threshold': 0.1698, # 'n_words': 67, 'n_windows': 1, 'warnings': [], 'confidence': 'high', # 'category': 'ai_generated'}] detector.predict(texts, operating_point="lenient") # flags 5% of human texts instead of 1% Detector.from_hub(segments=True).predict(texts) # also: verdict, ai_fraction, segmentsThe output shown is for a 67-word paragraph written by an AI assistant.
The command line works too: python detector.py "Some text.", with --segments,
--operating-point lenient and --dtype float16.
How detector.py scores a text, as the service does:
- Prepare. It applies the cleaning used on the training data, then removes invisible characters and folds homoglyphs.
- Score each window. The text is split into 512-token windows that overlap by 128 tokens. For each window the model gives a probability for each of 15 levels of AI share, from 0 (human) to 1 (AI-generated), and the window's score is the expected share.
- Combine. The window scores are averaged, as log-odds, into the text's score
(
probability). - Verdict and category. The score is compared with the threshold of the chosen operating point. A text below it is "human"; above it, "ai_assisted" up to a score of 0.75 and "ai_generated" beyond.
With segments=True, the service's mode, each sentence is also labelled. The text's verdict is
"ai" when at least 80% of its characters are labelled AI, "human" when at least 90% are labelled
human, otherwise "mixed". A long text's category comes from the score of its AI segments.
confidence (low / medium / high) says how far the evidence lies from the threshold. On
held-out data with as many human as AI texts, 0.3% of "ai / high" texts and 99.5% of
"human / high" texts are human (RESULTS.md).
Requirements:
- About 8.3 GB of GPU memory.
- The model runs in bfloat16 on GPUs that have it natively (Ampere and later) and in float32 on
older GPUs and the CPU.
dtype="float16"is several times faster on older GPUs and is what the service runs; on 774 test texts its scores were within 0.007 of float32's (RESULTS.md). - transformers 4.57 or later. 4.57.3, 5.14.1 and 5.23.0 give identical scores.
Without detector.py. For a single window, without the preparation (scores will differ a
little, more on texts with homoglyphs or invisible characters):
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
repo = "empirischtech/paratrace-v10"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo, dtype=torch.bfloat16).eval()
inputs = tokenizer("Text to check.", truncation=True, max_length=512, return_tensors="pt")
with torch.no_grad():
p = model(**inputs).logits.float().softmax(-1)[0] # 15 levels of the AI share
ai_share = (p * torch.linspace(0, 1, len(p))).sum().item() # 0 human ... 1 AI-generated
Training
Qwen3-4B-Base was fine-tuned as a classifier of the share of AI-written text: a head with 15 levels, read at a window's last token. It was trained in three steps, each a LoRA (rank 16 on all linear layers) merged into the weights:
- The main run: 2.59 million English texts, one epoch at learning rate 1e-4.
- A fine-tune on 499,000 texts: rows of the main run's data, more of RAID's training split, and new essays, academic texts and hard negatives with AI mirrors; one epoch at learning rate 5e-5.
- A short fine-tune on 114,000 texts: rows of the second step's data, the commercial humanizers' rewrites three times over, and polished human argument written before ChatGPT (newspaper opinion pieces, IELTS and TOEFL model answers by teachers) with AI mirrors; one epoch at learning rate 5e-5.
What the training data holds:
- Human texts from public corpora:
- reference and academic: Wikipedia, arXiv and peS2o abstracts and papers, PubMed;
- books and fiction: Project Gutenberg, CMU book summaries, WritingPrompts;
- online writing: Reddit, Stack Exchange, ELI5, blogs, Yelp reviews;
- news: XSum, CNN/DailyMail, CC-News, Global Voices;
- public records: patents, case law, UK Hansard, US government records;
- other: open textbooks, recipes, emails, speeches, and student essays (PERSUADE 2.0, ELLIPSE, ASAP 2.0, and the essays of English learners).
- AI "mirrors" of those human texts, each written on the same topic by one of more than 20 open-weight and commercial models, so the model learns authorship rather than topic.
- AI-assisted text: LLM paraphrases and rewrites of human texts. Each is labelled with an AI share found by aligning its clauses with the original: a reworded clause counts half, a new one in full (a paraphrase comes out at about 0.5).
- Humanized AI text: rewrites by ten commercial humanizer tools, by ParaTrace's own humanizer models, and by humanizing prompts.
- Other AI text: chat-assistant answers; the AI texts of HC3, MAGE, the COLING 2025 MGT shared task and part of RAID's training split.
- Hard negatives: human texts that earlier detectors wrongly flagged, each with an AI mirror.
- Held out for evaluation:
- Anthropic's persuasion essays;
- RAID's poetry domain, GPT-4, Mistral-chat and attacked texts;
- one of ParaTrace's humanizer models;
- the benchmark texts of UChicago, GEDE and Epoch AI.
Settings: 512-token windows, bfloat16, batch 64. Half of the texts longer than one window were cut at a random position, and a quarter of the fully human or fully AI texts cut to an excerpt of 15 to 60 words. Development AUROC: 0.9972.
Data. ParaTrace open data (CC BY-SA 4.0) holds 1.26 million of the 2.74 million distinct texts this model was trained on, its Wikipedia and arXiv paraphrases among them.
Intended use
- In scope:
- research on AI-generated text and its detection;
- evaluating and comparing detectors;
- one signal among others when a person reviews a text.
- Out of scope:
- using the output as the only basis for decisions about people (academic misconduct, hiring, publishing, moderation);
- non-English text;
- texts under 50 words;
- commercial use (see the license).
License
CC BY-NC-SA 4.0. The training data includes corpora licensed for non-commercial use only: PERSUADE 2.0 and ELLIPSE (CC BY-NC-SA 4.0), an English-learner essay corpus (CC BY-NC 4.0) and Yelp reviews (Yelp Dataset terms). The base model, Qwen3-4B-Base, is released by its authors under the Apache 2.0 License.
Citation
@misc{paratrace2026v10,
title = {ParaTrace v10},
author = {{ParaTrace}},
year = {2026},
howpublished = {\url{https://huggingface.co/empirischtech/paratrace-v10}}
}
@misc{qwen3technicalreport,
title = {Qwen3 Technical Report},
author = {{Qwen Team}},
year = {2025},
eprint = {2505.09388},
archivePrefix = {arXiv},
primaryClass = {cs.CL}
}
Contact
- Downloads last month
- -
Model tree for empirischtech/paratrace-v10
Base model
Qwen/Qwen3-4B-Base