syntaxsynth/tmmluplus

TMMLU+ : Large scale traditional chinese massive multitask language understanding

We present TMMLU+, a traditional Chinese massive multitask language understanding dataset. TMMLU+ is a multiple-choice question-answering dataset featuring 66 subjects, ranging from elementary to professional level.

The TMMLU+ dataset is six times larger and contains more balanced subjects compared to its predecessor, TMMLU. We have included benchmark results in TMMLU+ from closed-source models and 20 open-weight Chinese large language models, with parameters ranging from 1.8B to 72B. The benchmark results show that Traditional Chinese variants still lag behind those trained on major Simplified Chinese models.

Leaderboard ( Last Updated : 2025/04/06 )

Model	humanities	social sciences	STEM	Others	Average
openrouter/quasar-alpha	70.38	83.07	87.89	76.42	79.44
deepseek-chat	73.19	81.93	82.93	74.41	78.11
openrouter/optimus-alpha	67.73	78.68	84.14	74.73	76.32
Qwen/Qwen2.5-72B-Instruct-Turbo	67.59	79.36	82.57	72.65	75.54
gpt-4o-2024-08-06	65.48	78.23	81.39	71.24	74.08
claude-3-5-sonnet-20240620	73.23	78.27	68.50	69.35	72.34
gemini-2.0-flash-001	67.81	75.24	74.79	65.92	70.94
gemini-2.0-flash-lite-001	65.26	75.12	72.73	65.50	69.65
gemini-2.0-flash-lite-preview-02-05	64.66	73.48	70.00	63.90	68.01
qwen/qwen2.5-vl-32b-instruct	66.40	71.09	68.39	64.98	67.71
Qwen/QwQ-32B-Preview	57.98	70.94	72.87	63.59	66.35
claude-3-opus-20240229	60.34	70.12	67.43	62.32	65.05
gemini-1.5-pro	61.84	70.29	66.18	60.30	64.65
gpt-4o-mini-2024-07-18	55.01	67.09	73.16	61.36	64.15
mistralai/Mistral-Small-24B-Instruct-2501	54.56	68.32	73.25	59.25	63.85
meta-llama/llama-3.2-90b-vision-instruct	59.00	68.11	66.60	61.36	63.76
llama-3.1-70b-versatile	64.94	70.14	58.63	61.33	63.76
Qwen/Qwen2.5-7B-Instruct-Turbo	54.42	64.51	68.01	58.83	61.44
yentinglin/Llama-3-Taiwan-8B-Instruct	61.51	67.61	52.05	58.60	59.94
meta-llama/llama-4-scout	51.00	59.91	62.53	56.36	57.45
google/gemma-3-27b-it	52.54	57.69	58.59	50.50	54.83
claude-3-sonnet-20240229	52.06	59.38	49.87	51.64	53.24
Qwen2-7B-Instruct	55.66	66.40	27.18	55.32	51.14
meta-llama/llama-4-maverick	45.05	51.06	55.37	52.53	51.00
gemma2-9b-it	45.38	55.76	49.89	48.92	49.99
claude-3-haiku-20240307	47.48	54.48	48.47	48.77	49.80
gemini-1.5-flash	42.99	53.42	53.47	46.56	49.11
reka-flash	44.07	52.68	46.04	43.43	46.56
meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo	44.03	50.95	42.75	45.19	45.73
deepseek-reasoner	0.00	0.00	89.94	84.71	43.66
mixtral-8x7b-32768	44.75	50.34	32.60	43.76	42.86
meta-llama/Llama-3-70b-chat-hf	37.50	47.02	34.44	39.51	39.62
gemini-2.0-pro-exp-02-05	0.00	0.00	78.25	77.20	38.86
RekaAI/reka-flash-3	34.76	40.70	42.98	35.37	38.45
google/gemma-7b-it	34.00	35.70	31.89	33.79	33.84
reka-edge	31.84	39.40	30.02	32.36	33.41
meta-llama/Llama-3-8b-chat-hf	28.91	34.19	31.52	31.79	31.60
taide/Llama3-TAIDE-LX-8B-Chat-Alpha1	27.02	36.64	25.33	27.96	29.24
o1-preview-2024-09-12	0.00	0.00	81.55	0.00	20.39
claude-instant-1	0.00	0.00	25.24	20.30	11.39
Llama-3.1-8B-Instruct	0.00	0.00	15.53	0.00	3.88

How to use

from datasets import load_dataset
task_list = [
             'engineering_math', 'dentistry', 'traditional_chinese_medicine_clinical_medicine', 'clinical_psychology', 'technical', 'culinary_skills', 'mechanical', 'logic_reasoning', 'real_estate',
             'general_principles_of_law', 'finance_banking', 'anti_money_laundering', 'ttqav2', 'marketing_management', 'business_management', 'organic_chemistry', 'advance_chemistry',
             'physics', 'secondary_physics', 'human_behavior', 'national_protection', 'jce_humanities', 'politic_science', 'agriculture', 'official_document_management',
             'financial_analysis', 'pharmacy', 'educational_psychology', 'statistics_and_machine_learning', 'management_accounting', 'introduction_to_law', 'computer_science', 'veterinary_pathology',
             'accounting', 'fire_science', 'optometry', 'insurance_studies', 'pharmacology', 'taxation', 'trust_practice', 'geography_of_taiwan', 'physical_education', 'auditing', 'administrative_law',
             'education_(profession_level)', 'economics', 'veterinary_pharmacology', 'nautical_science', 'occupational_therapy_for_psychological_disorders',
             'basic_medical_science', 'macroeconomics', 'trade', 'chinese_language_and_literature', 'tve_design', 'junior_science_exam', 'junior_math_exam', 'junior_chinese_exam',
             'junior_social_studies', 'tve_mathematics', 'tve_chinese_language', 'tve_natural_sciences', 'junior_chemistry', 'music', 'education', 'three_principles_of_people',
             'taiwanese_hokkien'
            ]
for task in task_list:
  val = load_dataset('syntaxsynth/tmmluplus', task)['validation']
  dev = load_dataset('syntaxsynth/tmmluplus', task)['train']
  test = load_dataset('syntaxsynth/tmmluplus', task)['test']

Statistic on all four categories : STEM, Social Science, Humanities, Other

Category	Test	Dev	Validation
STEM	3458	70	385
Social Sciences	5958	90	665
Humanities	1763	35	197
Other (Business, Health, Misc.)	8939	135	995
Total	20118	330	2242

Citation

@inproceedings{tam2024tmmlu+,
  title={Tmmlu+: An improved traditional chinese evaluation suite for foundation models},
  author={Tam, Zhi Rui and Pai, Ya Ting and Lee, Yen-Wei and Shuai, Hong-Han and Chen, Jun-Da and Chu, Wei Min and Cheng, Sega},
  booktitle={First Conference on Language Modeling},
  year={2024}
}