Dataset Card for “ViANLI”

Dataset Summary

ViANLI (Vietnamese Adversarial Natural Language Inference) is the first adversarial benchmark dataset for Vietnamese NLI, designed to evaluate model robustness against complex linguistic phenomena. The dataset was constructed using a human-and-machine-in-the-loop approach with multi-round adversarial generation and dual human–machine verification. ViANLI contains over 10,000 high-quality premise–hypothesis pairs across 13 diverse domains from Vietnamese news articles.

Each pair is labeled as entailment, contradiction, or neutral, following the standard NLI framework. The dataset provides a challenging benchmark for evaluating reasoning robustness and supports research in both Vietnamese and multilingual NLI.

---

Languages

• Vietnamese (vi)

---

Dataset Structure

Data Instances

Each instance in ViANLI is a JSON line with the following fields:

Field	Type	Description
uid	string	Unique identifier for each instance
premise	string	The premise sentence extracted from Vietnamese news
hypothesis	string	The hypothesis sentence written by annotators
label	string	One of three classes: entailment, neutral, contradiction

Data Splits

Split	Size
train	8,012
validation	1,000
test	1,000

---

Dataset License

• License: CC BY-NC-SA 4.0 (Creative Commons Attribution–NonCommercial–ShareAlike 4.0 International License)

Citation Information

If you use this dataset, please cite:

@article{HUYNH2025130109,
title = {A New Benchmark Dataset and Mixture-of-Experts Language Models for Adversarial Natural Language Inference in Vietnamese},
journal = {Expert Systems with Applications},
pages = {130109},
year = {2025},
issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2025.130109},
url = {https://www.sciencedirect.com/science/article/pii/S095741742503725X},
author = {Tin Van Huynh, Kiet Van Nguyen and Ngan Luu-Thuy Nguyen},
}