mujian2026/multilingual-ai-text-detector

Multilingual AI Text Detector (ONNX)

English · 中文说明 · Source code · Live demo

Browser-ready ONNX derivatives of bibbbu/multilingual-ai-human-detector_xlm-roberta-base, exported at revision b4aa2e06d8428b4d27f2ef625267044d74b2eeb8 for local inference with Transformers.js and onnxruntime-web.

This model is a screening signal. It is not a forensic authorship test and must not be treated as proof that a person did or did not use generative AI.

At a glance

Decision question	Answer
Primary audience	Frontend engineers, ML engineers, researchers, and product teams that need browser-side text screening
Recommended variant	q4, about 181 MB; use q8 if the runtime cannot execute the q4 graph
Reference variant	fp32, about 1.11 GB
Output	Two-class classification: human and ai
Evidence languages	English, Simplified Chinese, and Vietnamese in the upstream evaluation
Experimental website languages	English, Simplified Chinese, Traditional Chinese, and Japanese
Privacy model	The demo performs inference locally after public model files are downloaded
License	MIT for the published conversion project; retain upstream model attribution

Who should use this model?

Browser and frontend developers

Use it when you need a JavaScript-callable ONNX model, want text to remain in the browser, or need a smaller q4/q8 alternative to a server-only detector. The important trade-offs are download size, first-load time, browser memory, operator support, and a reliable fallback path.

ML engineers and researchers

Use it as a reproducible browser packaging of an existing multilingual classifier. The source revision, conversion script, quantizers, fixtures, and report are available in the companion GitHub repository. The upstream benchmark and its small, in-domain test set should be reviewed before drawing conclusions.

Product teams

Use it for a low-friction first-pass signal in writing, education, moderation, or content-quality workflows. It can reduce server-side inference work and keep ordinary input text out of an application API, but the result should be combined with writing history, source review, or human judgment.

Do not use this model as the sole basis for academic misconduct decisions, employment decisions, legal conclusions, or irreversible moderation actions.

Variant and deployment guide

Variant	Approx. size	Best use	Trade-off
q4	181 MB	Default browser download	Smallest file; requires q4 operator support
q8	279 MB	Browser fallback	Larger file; broader quantized-runtime compatibility
fp32	1.11 GB	Reference comparison or special deployment	Largest download and memory cost

Recommended production chain:

1. Load q4 from this Hugging Face repository.
2. Fall back to q8 only when q4 loading or execution is unsupported.
3. Use an explicitly configured R2/CDN mirror only when the primary URL fails.
4. Pin a model revision and cache immutable files; do not mix q4, q8, and FP32 under one unversioned path.

Installation

Install Transformers.js in a browser project:

npm install @huggingface/transformers

The conversion scripts and Python environment are maintained separately in the GitHub source repository. You do not need Python to consume the published browser model.

Usage

Load q4 with Transformers.js:

import { pipeline } from '@huggingface/transformers'

const detector = await pipeline(
  'text-classification',
  'mujian2026/multilingual-ai-text-detector',
  { dtype: 'q4' },
)

const result = await detector('This is a sample passage.')
console.log(result)

The returned label is human or ai. The score is a model confidence signal, not a calibrated probability of authorship. In a multilingual product, select the UI language for user experience, but do not present the result as a language-independent guarantee.

Benchmark

The upstream model card reports F1 on 270 held-out examples, with 90 examples per language:

Language	Reported F1
English	0.9890
Vietnamese	0.9783
Chinese	0.9462
Overall	0.9710

These are in-domain results from the upstream test split. They are not a universal accuracy claim. On the same reported split, a TF-IDF logistic regression baseline reached 0.9776 overall F1, while the XLM-R detector reached 0.9710. This model should therefore be evaluated on the domain, languages, and generators that matter to your application before adoption.

The source evaluation used AI answers generated with Qwen2.5-1.5B-Instruct. Performance on GPT, Gemini, Claude, newer Qwen models, translated text, paraphrases, or heavily edited text is not established by that benchmark.

Quantization benchmark

The published variants were compared with the FP32 ONNX export on six short English, Chinese, and Japanese fixtures using CPU ONNX Runtime:

Variant	Approx. size	Mean probability delta	Max delta	Label flips
FP32	1.11 GB	reference	reference	reference
q4	181 MB	0.0041	0.0231	0 / 6
q8	279 MB	0.0804	0.3672	0 / 6

All three variants were also loaded and executed with the browser WASM runtime before publication. This is a conversion smoke test, not a representative multilingual benchmark. Re-run evaluation after changing the source revision, quantizer, or runtime.

Model origin and training data

This repository contains an ONNX export and quantized derivatives; it does not claim to have trained a new detector. The source checkpoint is fine-tuned from FacebookAI/xlm-roberta-base and predicts human versus ai.

According to the upstream model card:

• Training used 900 QA pairs: 300 English, 300 Chinese, and 300 Vietnamese.
• Human examples came from HC3 Reddit ELI5, HC3-Chinese Open QA, and Vietnamese Reddit communities.
• AI examples were generated with Qwen2.5-1.5B-Instruct.
• The setup used three epochs, learning rate 2e-5, maximum length 256, and a prompt-level 70/15/15 train/validation/test split with seed 42.
• The model has approximately 279M XLM-RoBERTa-base parameters plus a sequence-classification head.

Language scope and limitations

The upstream evaluation covers English, Simplified Chinese, and Vietnamese. The MyToolster website additionally routes Traditional Chinese and Japanese through the same XLM-R classifier. Traditional Chinese and Japanese are experimental signals here and are not independently calibrated probabilities.

The classifier learns patterns from its training distribution; it does not inspect hidden metadata or a universal AI signature. False positives and false negatives are expected. Short inputs, translation, editing, paraphrasing, domain vocabulary, and generator changes can materially affect the result.

Use the detector as one input to a broader review, not as a standalone verdict.

Files

q4/onnx/model_q4.onnx             # 4-bit browser variant
q8/onnx/model_quantized.onnx      # 8-bit browser variant
fp32/onnx/model.onnx              # FP32 reference variant
*/config.json                     # label mapping and model configuration
*/tokenizer.json                  # tokenizer files for each variant
quantization-report.json          # conversion comparison report

The conversion and quantization source is maintained in the GitHub repository.

Citation and license

If you use the underlying detector or cite its training/evaluation results, cite the upstream work:

@misc{vu2024multilingual,
  title  = {Multilingual AI-Human Text Detection},
  author = {Vu, Tuong Vy},
  year   = {2024},
  url    = {https://github.com/vutuongvy101/multilingual-ai-human-text-detection}
}

The published conversion project is MIT-licensed. The fine-tuned source checkpoint and training work belong to the upstream authors; retain their attribution and license when redistributing the model files.