chuxuecao/HKMMLU

HKMMLU Dataset

| 📖 Paper |

HKMMLU dataset is a multi-task language understanding dataset that evaluates both Hong Kong linguistic competence and socio-cultural knowledge.

Dataset Summary

The HKMMLU includes 26,698 multi-choice questions across 66 subjects, organized into four categories: Science, Technology, Engineering, and Mathematics (STEM), Social Sciences, Humanities, and Other. To evaluate the multilingual understanding ability of LLMs, it additionally comprises 88,292 Mandarin-Cantonese translation tasks.

🏆 Mini-Leaderboard

Model performance on multi-choice tasks in HKMMLU using zero-shot prompting.

Model	HKMMLU
DeepSeek-V3	74.8
GPT-4o	70.3
Qwen2.5-72B-Instruct	69.1
Llama-3-Taiwan-70B-Instruct	67.4
Claude 3.7 Sonnet	66.7
Qwen2.5-14B-Instruct	62.3
Yi-1.5-34B-Chat	60.6
Mistral-Large-Instruct	60.0
Llama-3-70B-Instruct	59.2
Llama-3.1-70B-Instruct	58.7
Gemma-2-27B-IT	57.4
Qwen2.5-7B-Instruct	57.2
Yi-1.5-9B-Chat	52.6
Llama-3-Taiwan-8B-Instruct	51.9
Qwen2.5-3B-Instruct	50.3
GLM-4-9B-Chat	49.0
Mistral-Small-Instruct	45.2
Llama-3-8B-Instruct	45.1
Gemma-2-2B-IT	42.0
Llama-3.1-8B-Instruct	40.0

Updates

2025-05-08: Update translation tasks to 88,292 for quality enhancement.

Load data

from datasets import load_dataset
hkmmlu_hklaw = load_dataset(r"zhenghaozhu/HKMMLU", 'hk_law')
print(hkmmlu_hklaw)
print(hkmmlu_hklaw['test'][0])

Output

DatasetDict({
    test: Dataset({
        features: ['Question', 'A', 'B', 'C', 'D', 'Answer'],
        num_rows: 407
    })
    validation: Dataset({
        features: ['Question', 'A', 'B', 'C', 'D', 'Answer'],
        num_rows: 5
    })
})
{'Question': '以下哪一選項最準確描述制定《中華人民共和國香港特別行政區維護國家安全法》的法律依據？', 'A': '《中華人民共和國憲法》和《中華人民共和國香港特別行政區基本法》', 'B': '《中華人民共和國憲法》、《中華人民共和國香港特別行政區基本法》和《全國人民代表大會關於建立健全香港特別行政區維護國家安全的法律制度和執行機制的決定》', 'C': '《中華人民共和國憲法》、《中華人民共和國香港特別行政區基本法》和《中華人民共和國刑法》', 'D': '《全國人民代表大會關於建立健全香港特別行政區維護國家安全的法律制度和執行機制的決定》、《中華人民共和國刑法》和《中華人民共和國國家安全法》', 'Answer': 'B'}

Load all data at once

from datasets import load_dataset

multi_choice_task_list = ['basic_medical_science', 'business_management', 'chinese_conditions', 'chinese_history', 'chinese_world_culture', 'chinese_world_geography', 'clinical_psychology', 'college_math', 'culture_commonsense', 'dentistry', 'economics', 'educational_psychology', 'financial_analysis', 'fire_science', 'hk_celebrity', 'hk_cultural_art', 'hk_current_affairs', 'hk_education', 'hk_entertainment_industry_management', 'hk_events_festivals', 'hk_film_television', 'hk_government_political_system', 'hk_historical_development_geography', 'hk_history', 'hk_human_geography', 'hk_law', 'hk_life_knowledge', 'hk_literature', 'hk_livelihood', 'hk_music', 'hk_party_politics', 'hk_physical_geography', 'hk_politicians_events', 'hk_society', 'hk_transport_infrastructure', 'hk_transportation', 'hk_urban_regional_development', 'hkdse_biology', 'hkdse_business_accounting_finance', 'hkdse_chemistry', 'hkdse_economics', 'hkdse_geography', 'hkdse_information_and_communication_technology', 'hkdse_mathematics', 'hkdse_physics', 'hkdse_tourism', 'human_behavior', 'insurance_studies', 'logic_reasoning', 'management_accounting', 'marketing_management', 'mechanical', 'nautical_science', 'optometry', 'organic_chemistry', 'pharmacology', 'pharmacy', 'physics', 'science_technology_commonsense', 'social_commonsense', 'statistics_and_machine_learning', 'technical', 'traditional_chinese_medicine_clinical_medicine', 'veterinary_pathology', 'veterinary_pharmacology', 'world_entertainment']
translate_task_list = ['Mandarin_To_Cantonese', 'Cantonese_To_Mandarin']

hkmmlu_multichoice = {k: load_dataset(r"zhenghaozhu/HKMMLU", k) for k in multi_choice_task_list}
hkmmlu_translate = {k: load_dataset(r"zhenghaozhu/HKMMLU", k) for k in translate_task_list}

Citation

@misc{cao2025measuringhongkongmassive,
      title={Measuring Hong Kong Massive Multi-Task Language Understanding}, 
      author={Chuxue Cao and Zhenghao Zhu and Junqi Zhu and Guoying Lu and Siyu Peng and Juntao Dai and Weijie Shi and Sirui Han and Yike Guo},
      year={2025},
      eprint={2505.02177},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.02177}, 
}