metadata
license: cc-by-nc-nd-4.0
task_categories:
- text-classification
language:
- en
tags:
- agent
pretty_name: ReaMent
size_categories:
- 1M<n<10M
Boosting Large Language Models for Mental Manipulation Detection via Data Augmentation and Distillation
✨ Like ReaMent? Give us a ⭐ Star on GitHub! Your support keeps us going! Yuansheng-Gao/MentalMAD
🌿 ReaMent Dataset Card
A multi-round, real-world conversation-based mental manipulation detection dataset.
🧠 Dataset Summary
The ReaMent dataset was created to address the lack of real-world data in the field of mental manipulation detection.
- Source: The dataset is built from the YTD-18M corpus, which contains over 18 million dialogue-like segments extracted from unscripted interactions in web videos. These dialogues cover a wide range of everyday scenarios, such as interviews, group discussions, and situational conversations.
- Size: The final dataset consists of 5,000 high-quality annotated dialogues.
- Diversity: ReaMent captures a broader range of conversational contexts compared to scripted data, providing more natural and spontaneous interaction patterns.
- Statistics: Around 68.3% of dialogues in ReaMent were labeled as containing mental manipulation, while 31.7% were labeled as non-manipulative. The dataset has an average of 4 dialogue turns and 80 words per dialogue.
🤗 Key Contributions
- Real-World Representation: Unlike scripted or domain-specific datasets (e.g., MentalManip and LegalCon), ReaMent captures natural dialogues, making it valuable for detecting real-world mental manipulation.
- Scalability: It complements smaller datasets, offering richer and more representative data for training models that aim to detect manipulative behaviors in social interactions.
💻 Usage
from datasets import load_dataset
ds = load_dataset("YSGao/ReaMent")
📝 Citation
@misc{gao2026boostinglargelanguagemodels,
title={Boosting Large Language Models for Mental Manipulation Detection via Data Augmentation and Distillation},
author={Yuansheng Gao and Peng Gao and Han Bao and Bin Li and Jixiang Luo and Zonghui Wang and Wenzhi Chen},
year={2026},
eprint={2505.15255},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2505.15255},
}