Multiscale Positive-Unlabeled Detection of AI-Generated Texts

July 24, 2025 · View on GitHub

This is the official repo of our ICLR'24 Spotlight paper "Multiscale Positive-Unlabeled Detection of AI-Generated Texts".

News

7/24/2025: The latest v3 models and demos are released on ModelScope! DEMO

6/25/2025 🔥🔥 MAJOR UPDATE: We release advanced (v3) detectors targeting latest LLMs, including DeepSeek-V3, GPT-4, and reasoning models. En-v3 / En-v3-short / Zh-v3 / Zh-v3-short / DEMO

6/1/2025: We release a beta version targeting latest LLMs, including DeepSeek-V3 and GPT-4. En-beta Zh-beta

3/25/2025: We release a demo (with both English and Chinese) on HuggingFace.

3/6/2025: We will update a brand-new detector version to align with the latest LLMs. Please keep tuned!

BibTex formatted citation:

@misc{tian2023multiscale,
      title={Multiscale Positive-Unlabeled Detection of AI-Generated Texts}, 
      author={Yuchuan Tian and Hanting Chen and Xutao Wang and Zheyuan Bai and Qinghua Zhang and Ruifeng Li and Chao Xu and Yunhe Wang},
      year={2023},
      eprint={2305.18149},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Detector Models

We have open-sourced detector models in the paper as follows.

Links for Detectors: Google Drive Baidu Disk (PIN:1234)

We have also uploaded detector models to HuggingFace, where easy-to-use DEMOs and online APIs are provided.

VariantsHC3-Full-EnHC3-Sent-En
seed098.6882.84
seed1 HuggingFace: en v198.5687.06
seed297.9786.02
Avg.98.40\pm\0.3185.31\pm\1.80

Stronger Detectors

We have also open-sourced detector models with strengthened training strategies. Specifically, we develop a strong Chinese detector AIGC_detector_zhv2, which demonstrates similar performance to SOTA closed-source Chinese detectors on various texts, including news articles, poetry, essays, etc. The DEMOs and APIs are available on HuggingFace.

DetectorGoogle DriveBaidu DiskHuggingFace Link
English, version 2 (env2)Google DriveBaidu Disk (PIN:1234)en v2
Chinese, version 2 (zhv2)Google DriveBaidu Disk (PIN:1234)zh v2

About the Dataset

Here we provide the official link for the HC3 dataset: Dataset Link. We also provide identical dataset copies on Google Drive and Baidu Disk (PIN:1234) for your ease of use. We acknowledge the marvelous work by HC3 authors.

Data Preprocessing

In Appendix B of our paper, we proposed the removal of redundant spaces in human texts of the HC3-English dataset. We have provided a helper function en_cleaning in corpus_cleaning_kit.py that takes a sentence string as input and returns a preprocessed sentence without redundant spaces.

Here we provide a cleaned version of HC3-English. In this version, all answers are cleaned (i. e. redundant spaces are removed). However, please use the original version of HC3 for all experiments in our paper, as we have embedded the cleaning procedures in the training & validation scripts.

CLEANED HC3-English Link: Google Drive Baidu Disk (PIN:1234)

Preparation

  • Install requirement packages:
pip install -r requirements.txt

  • Download datasets to directory: ./data

  • Download nltk package punct (This step could be done by nltk api: nltk.download('punkt'))

  • Download pretrained models (This step could be automatically done by transformers)

Before running, the directory should contain the following files:

├── data
│   ├── unfilter_full
│   │   ├── en_test.csv
│   │   └── en_train.csv
│   └── unfilter_sent
│       ├── en_test.csv
│       └── en_train.csv
├── README.md
├── corpus_cleaning_kit.py
├── dataset.py
├── multiscale_kit.py
├── option.py
├── pu_loss_mod.py
├── prior_kit.py
├── requirements.txt
├── train.py
└── utils.py

Training

The script for training is train.py.

RoBERTa on HC3-English

Commands for seed=0,1,2:

CUDA_VISIBLE_DEVICES=0 python train.py --batch-size 32 --max-sequence-length 512 --train-data-file unfilter_full/en_train.csv --val-data-file unfilter_full/en_test.csv --model-name roberta-base --local-data data --lamb 0.4 --prior 0.2 --pu_type dual_softmax_dyn_dtrun --len_thres 55 --aug_min_length 1 --max-epochs 1 --weight-decay 0 --mode original_single --aug_mode sentence_deletion-0.25 --clean 1 --val_file1 unfilter_sent/en_test.csv --quick_val 1 --learning-rate 5e-05 --seed 0

CUDA_VISIBLE_DEVICES=0 python train.py --batch-size 32 --max-sequence-length 512 --train-data-file unfilter_full/en_train.csv --val-data-file unfilter_full/en_test.csv --model-name roberta-base --local-data data --lamb 0.4 --prior 0.2 --pu_type dual_softmax_dyn_dtrun --len_thres 55 --aug_min_length 1 --max-epochs 1 --weight-decay 0 --mode original_single --aug_mode sentence_deletion-0.25 --clean 1 --val_file1 unfilter_sent/en_test.csv --quick_val 1 --learning-rate 5e-05 --seed 1

CUDA_VISIBLE_DEVICES=0 python train.py --batch-size 32 --max-sequence-length 512 --train-data-file unfilter_full/en_train.csv --val-data-file unfilter_full/en_test.csv --model-name roberta-base --local-data data --lamb 0.4 --prior 0.2 --pu_type dual_softmax_dyn_dtrun --len_thres 55 --aug_min_length 1 --max-epochs 1 --weight-decay 0 --mode original_single --aug_mode sentence_deletion-0.25 --clean 1 --val_file1 unfilter_sent/en_test.csv --quick_val 1 --learning-rate 5e-05 --seed 2

BERT on HC3-English

Commands for seed=0,1,2:

CUDA_VISIBLE_DEVICES=0 python train.py --batch-size 32 --max-sequence-length 512 --train-data-file unfilter_full/en_train.csv --val-data-file unfilter_full/en_test.csv --model-name bert-base-cased --local-data data --lamb 0.5 --prior 0.3 --pu_type dual_softmax_dyn_dtrun --len_thres 60 --aug_min_length 1 --max-epochs 1 --weight-decay 0 --mode original_single --aug_mode sentence_deletion-0.25 --clean 1 --val_file1 unfilter_sent/en_test.csv --quick_val 1 --learning-rate 5e-05 --seed 0


CUDA_VISIBLE_DEVICES=0 python train.py --batch-size 32 --max-sequence-length 512 --train-data-file unfilter_full/en_train.csv --val-data-file unfilter_full/en_test.csv --model-name bert-base-cased --local-data data --lamb 0.5 --prior 0.3 --pu_type dual_softmax_dyn_dtrun --len_thres 60 --aug_min_length 1 --max-epochs 1 --weight-decay 0 --mode original_single --aug_mode sentence_deletion-0.25 --clean 1 --val_file1 unfilter_sent/en_test.csv --quick_val 1 --learning-rate 5e-05 --seed 1


CUDA_VISIBLE_DEVICES=0 python train.py --batch-size 32 --max-sequence-length 512 --train-data-file unfilter_full/en_train.csv --val-data-file unfilter_full/en_test.csv --model-name bert-base-cased --local-data data --lamb 0.5 --prior 0.3 --pu_type dual_softmax_dyn_dtrun --len_thres 60 --aug_min_length 1 --max-epochs 1 --weight-decay 0 --mode original_single --aug_mode sentence_deletion-0.25 --clean 1 --val_file1 unfilter_sent/en_test.csv --quick_val 1 --learning-rate 5e-05 --seed 2

Updates

A major update (v3) is ongoing, with updated English and Chinese detectors, datasets, and codes. This time, our detector targets at corpora from MiniMax-Text-01, Llama-3.3-70B-Instruct, Qwen2.5-72B-Instruct, DeepSeek-V3, and GPT-4-Turbo. Besides, Chain-of-Thought (CoT) models are covered, including DeepSeek-R1, GPT-o3, and QwQ-32B. Please keep tuned for future updates!

VersionEnglishChineseComments
6/25/2025 AIGC-Detector-V3En_v3Zh_v3Version 3. Strongest performance on normal AI texts. CoT LLMs are covered.
6/25/2025 AIGC-Detector-V3-ShortEn_v3_shortZh_v3_shortVersion 3 targeting shorter texts without sacrificing longer ones. CoT LLMs are covered.
5/31/2025 AIGC-Detector-BetaEn_betaZh_betaA beta version. Stronger versions will be released in the next update.

Acknowledgement

Our code refers to the following GitHub repo:

https://github.com/openai/gpt-2-output-dataset

We sincerely thank their authors for open-sourcing.