ikala/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.

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('ikala/tmmluplus', task)['validation']
  dev = load_dataset('ikala/tmmluplus', task)['train']
  test = load_dataset('ikala/tmmluplus', task)['test']

For each dataset split

for row in test:
  print(row)
  break
>> Dataset({
    features: ['question', 'A', 'B', 'C', 'D', 'answer'],
    num_rows: 11
})

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

Benchmark on direct prompting

model	STEM	Social Science	Humanities	Other	Average
Qwen/Qwen-72B	61.12	71.65	63.00	61.31	64.27
gpt-4-0613	60.36	67.36	56.03	57.62	60.34
Qwen-max	59.92	66.95	57.43	56.48	60.20
Qwen/Qwen-72B-Chat	55.15	66.20	55.65	57.19	58.55
Qwen/Qwen-14B	46.94	56.69	49.43	48.81	50.47
Gemini-pro	45.38	57.29	48.80	48.21	49.92
01-ai/Yi-34B-Chat	40.24	56.77	53.99	47.58	49.64
Reka Flash	45.26	52.91	46.31	43.76	47.06
Qwen/Qwen-14B-Chat	43.86	53.29	44.78	45.13	46.77
01-ai/Yi-6B-Chat	39.62	50.24	44.44	44.26	44.64
Claude-1.3	42.65	49.33	42.16	44.14	44.57
gpt-3.5-turbo-0613	41.56	46.72	36.73	42.03	41.76
CausalLM/14B	39.83	44.50	39.61	41.97	41.48
Skywork/Skywork-13B-base	36.93	47.27	41.04	40.10	41.33
Qwen/Qwen-7B	37.53	45.48	38.09	38.96	40.01
Qwen/Qwen-7B-Chat	33.32	44.64	40.27	39.89	39.53
vivo-ai/BlueLM-7B-Base	33.94	41.52	37.38	38.74	37.90
baichuan-inc/Baichuan2-13B-Chat	29.64	43.73	37.36	39.88	37.65
Qwen/Qwen-1_8B	32.65	38.95	38.34	35.27	36.30
Claude-2	39.65	39.09	28.59	37.47	36.20
THUDM/chatglm3-6b	31.05	39.31	35.64	35.60	35.40
deepseek-ai/deepseek-llm-7b-chat	29.82	42.29	34.24	34.31	35.17
CausalLM/7B	31.03	38.17	35.87	35.39	35.11
Azure99/blossom-v3_1-mistral-7b	32.80	36.91	32.36	34.53	34.15
Reka Edge	30.02	39.40	31.84	32.36	33.41
microsoft/Orca-2-13b	24.69	39.18	33.60	31.99	32.37
Qwen/Qwen-1_8B-Chat	26.60	36.36	31.81	31.96	31.68
TigerResearch/tigerbot-13b-chat-v3	24.73	29.63	25.72	27.22	26.82
hongyin/mistral-7b-80k	24.26	23.76	22.56	24.57	23.79
deepseek-ai/deepseek-llm-67b-chat	19.10	26.06	21.51	21.77	22.11
yentinglin/Taiwan-LLM-13B-v2.0-chat	18.53	27.65	17.77	21.49	21.36
GeneZC/MiniChat-3B	17.66	23.35	22.71	20.34	21.02
LinkSoul/Chinese-Llama-2-7b	16.55	18.39	12.97	16.13	16.01
yentinglin/Taiwan-LLM-7B-v2.1-chat	14.99	16.23	15.00	16.22	15.61
Claude-instant-1	12.52	17.13	15.10	13.57	14.58
FlagAlpha/Atom-7B	5.60	13.57	7.71	11.84	9.68

Results via ievals ( settings : 0-shot direct answering )

Citation

@article{ikala2024improved,
  title={An Improved Traditional Chinese Evaluation Suite for Foundation Model},
  author={Tam, Zhi-Rui and Pai, Ya-Ting and Lee, Yen-Wei and Cheng, Sega and Shuai, Hong-Han},
  journal={arXiv preprint arXiv:2403.01858},
  year={2024}
}