nguyenthanhasia/vbench-dataset

V-Bench: Vietnamese LLM Benchmark Dataset

Dataset Summary

V-Bench is a comprehensive benchmark dataset for evaluating Vietnamese Large Language Models (LLMs). It provides a rigorous, multi-dimensional evaluation framework with 5 specialized data groups validated by 15+ expert reviewers. The dataset assesses both knowledge capabilities and agentic behaviors of Vietnamese LLMs across 9+ academic domains, safety and sovereignty, culture, and healthcare.

Dataset Details

• Total Samples: 9,141
• Total Size: 5.24 MB
• Language: Vietnamese (vi)
• License: Unknown
• Organization: Center for AI Research (CAIR), VinUniversity
• Contact: vbench-support@vinuni.edu.vn
• Website: https://vbench.ai

Dataset Structure

The dataset contains the following columns:

Column	Type	Description
id	int64	Unique identifier for each question
question	string	The Vietnamese question text
choices	string	Multiple choice options (if applicable)
function	string	Function calling specification (if applicable)
domain	string	Domain category of the question

Domain Distribution

Domain	Samples
hatespeech	2,000
politics_easy	1,000
politics_advanced	1,000
agentic	1,000
psychology	798
literature	601
dialect	562
medicine	490
chemistry	334
philosophy	263
logics	225
culture	217
laws	191
computer_science	188
physics	147
mathematics	125

Features

• Multi-domain Coverage: 16 different domains covering academic knowledge, safety, culture, and healthcare
• Expert Validation: All questions validated by 15+ expert reviewers using blind review methodology
• Multiple Task Types: Supports both multiple-choice QA and function-calling tasks
• Vietnamese Language: Specifically designed for Vietnamese LLM evaluation
• Comprehensive Metrics: Includes accuracy and function-call-correctness evaluation metrics

Usage

Loading the Dataset

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset('nguyenthanhasia/vbench-dataset')

# Access the train split
train_data = dataset['train']

# Get a specific sample
sample = train_data[0]
print(sample['question'])
print(sample['domain'])

Filtering by Domain

from datasets import load_dataset

dataset = load_dataset('nguyenthanhasia/vbench-dataset')
train_data = dataset['train']

# Filter by domain
hatespeech_data = train_data.filter(lambda x: x['domain'] == 'hatespeech')
politics_data = train_data.filter(lambda x: x['domain'] in ['politics_easy', 'politics_advanced'])

Dataset Splits

The dataset is provided as a single train split containing all 9,141 samples. For evaluation and scoring, please visit https://vbench.ai/submission.

Evaluation Metrics

• Accuracy: Percentage of correct answers
• Function-Call-Correctness: Percentage of correctly executed function calls

Domain Categories

The dataset covers the following 16 domains:

1. hatespeech: Detection and analysis of hateful content
2. politics_easy: Basic political knowledge questions
3. politics_advanced: Advanced political analysis and reasoning
4. agentic: Agent-based task execution and reasoning
5. psychology: Psychological concepts and theories
6. literature: Vietnamese and world literature
7. dialect: Vietnamese dialects and regional language variations
8. medicine: Medical knowledge and healthcare
9. chemistry: Chemistry concepts and reactions
10. philosophy: Philosophical concepts and theories
11. logics: Logical reasoning and problem-solving
12. culture: Vietnamese culture and traditions
13. laws: Vietnamese laws and legal systems
14. computer_science: Computer science and programming
15. physics: Physics concepts and principles
16. mathematics: Mathematical concepts and calculations

Citation

@dataset{vbench2026,
  title={V-Bench: Vietnamese LLM Benchmark Dataset},
  organization={Center for AI Research (CAIR), VinUniversity},
  year={2026},
  url={https://vbench.ai},
  version={v2026.07.15}
}

Contact & Support

• Email: vbench-support@vinuni.edu.vn
• Website: https://vbench.ai
• Submission Portal: https://vbench.ai/submission

Version History

• v2026.07.15: Optimized for Hugging Face Dataset Viewer compatibility
• v2026.03.28: Initial public release with 16 domains and 9,141 samples

License

The dataset is provided under an unknown/custom license. Please refer to the official V-Bench website for licensing details.

Disclaimer

This dataset is intended for research and benchmarking purposes. Users are responsible for ensuring compliance with all applicable laws and regulations when using this dataset.