metadata

license: mit
task_categories:
  - question-answering
language:
  - zh
tags:
  - medical
  - tcm
  - traditional Chinese medicine
  - eval
  - benchmark
  - test

Description

This dataset can be used to evaluate the capabilities of large language models in traditional Chinese medicine and contains multiple-choice, multiple-answer, and true/false questions.

Changelog

2024-08-28: Added 7226 questions.
2024-08-09: The benchmark code is available at https://github.com/huangxinping/HWTCMBench.
2024-08-02: System prompts are removed to ensure the purity of the evaluation results.
2024-07-20: Debut.

Examples

multiple-answers questions（多选题）

[
  {
    "instruction": "便秘的预防调护应注意\nA.保持心情舒畅\nB.少吃辛辣刺激性食物\nC.适当摄入油脂\nD.积极治疗肛门直肠疾病\nE.按时登厕",
    "input": "",
    "output": "ABCDE"
  }
]

multiple-choice questions（单选题）

[
  {
    "instruction": "患者，男，50岁。眩晕欲仆，头摇而痛，项强肢颤，腰膝疫软，舌红苔薄白，脉弦有力。其病机是\nA．肝阳上亢\nB．肝肾阴虚\nC．肝阳化风\nD．阴虚风动\nE．肝血不足",
    "input": "",
    "output": "C"
  }
]

True/False questions（判断题）

[
  {
    "instruction": "秦医医和提出了“六气病源说”。",
    "input": "",
    "output": "正确"
  },
  {
    "instruction": "中风中经络邪盛时也可出现神志改变",
    "input": "",
    "output": "错误"
  }
]

Evaluation

	multiple-choice questions	multiple-answers questions	True/False questions
llama3:8b	21.94%	17.71%	46.56%
phi3:14b-instruct	26.93%	1.04%	38.93%
aya:8b	17.85%	1.04%	34.35%
mistral:7b-instruct	21.76%	2.08%	48.09%
qwen1.5-7b-chat	51.35%	13.54%	46.56%
qwen1.5-14b-chat	69.94%	78.12%	31.30%
huangdi-13b-chat	21.73%	45.83%	0.00%
canggong-14b-chat(SFT) Ours	55.98%	4.17%	23.66%
canggong-14b-chat(DPO) Ours	72.33%	2.08%	45.80%

Canggong-14b-chat is an LLM of traditional Chinese medicine still in training.

Citation

If you find this project useful in your research, please consider cite:

@misc{hwtcm2024,
    title={{hwtcm} a traditional Chinese medicine QA dataset for evaluating large language models},
    author={Haiwei AI Team},
    howpublished = {\url{https://huggingface.co/datasets/Monor/hwtcm}},
    year={2024}
}