Multilingual Medicine: Model, Dataset, Benchmark, Code

Covering English, Chinese, French, Hindi, Spanish, Hindi, Arabic So far

📃 Paper • 🌐 Demo • 🤗 ApolloCorpus • 🤗 XMedBench
中文 | English

🌈 Update

[2024.04.25] MedJamba released, train and evaluation code refer to repo.
[2024.03.07] Paper released.
[2024.02.12] ApolloCorpus and XMedBench is published！🎉
[2024.01.23] Apollo repo is published！🎉

Results

🤗 Apollo-0.5B • 🤗 Apollo-1.8B • 🤗 Apollo-2B • 🤗 Apollo-6B • 🤗 Apollo-7B • 🤗 Apollo-34B • 🤗 Apollo-72B

🤗 MedJamba

🤗 Apollo-0.5B-GGUF • 🤗 Apollo-2B-GGUF • 🤗 Apollo-6B-GGUF • 🤗 Apollo-7B-GGUF

Usage Format

0.5B, 1.8B, 2B, 6B, 7B: User:{query}\nAssistant:{response}<|endoftext|>
34B, 72B: <|User|>:{query}\n<|Assistant|>:{response}<|endoftext|>

Dataset & Evaluation

Dataset 🤗 ApolloCorpus
Click to expand
- Zip File
- Data category
  - Pretrain:
    - data item:
      
      json_name: {data_source}{language}{data_type}.json
      
      data_type: medicalBook, medicalGuideline, medicalPaper, medicalWeb(from online forum), medicalWiki
      
      language: en(English), zh(chinese), es(spanish), fr(french), hi(Hindi)
      
      data_type: qa(generated qa from text)
      
      data_type==text: list of string
      [ "string1", "string2", ... ]
      
      data_type==qa: list of qa pairs(list of string)
      [ [ "q1", "a1", "q2", "a2", ... ], ... ]
  - SFT:
    - json_name: {data_source}_{language}.json
    - data_type: code, general, math, medicalExam, medicalPatient
    - data item: list of qa pairs(list of string)
      [ [ "q1", "a1", "q2", "a2", ... ], ... ]
Evaluation 🤗 XMedBench
Click to expand
- EN:
  - MedQA-USMLE
  - MedMCQA
  - PubMedQA: Because the results fluctuated too much, they were not used in the paper.
  - MMLU-Medical
    - Clinical knowledge, Medical genetics, Anatomy, Professional medicine, College biology, College medicine
- ZH:
  - MedQA-MCMLE
  - CMB-single: Not used in the paper
    - Randomly sample 2,000 multiple-choice questions with single answer.
  - CMMLU-Medical
    - Anatomy, Clinical_knowledge, College_medicine, Genetics, Nutrition, Traditional_chinese_medicine, Virology
  - CExam: Not used in the paper
    - Randomly sample 2,000 multiple-choice questions
- ES: Head_qa
- FR: Frenchmedmcqa
- HI: MMLU_HI
  - Clinical knowledge, Medical genetics, Anatomy, Professional medicine, College biology, College medicine
- AR: MMLU_Ara
  - Clinical knowledge, Medical genetics, Anatomy, Professional medicine, College biology, College medicine

Results reproduction

Click to expand

We take Gemma-2b as example

Download Dataset for project:
```
bash 0.download_data.sh
```
Prepare test and dev for specific model:
- Create test data for with special token, you can use ./util/check.ipynb to check models' special tokens
```
bash 1.data_process_test&dev.sh
```
Prepare train data for specific model (Create tokenized data in advance):
- You can adjust data Training order and Training Epoch in this step
```
bash 2.data_process_train.sh
```
Train the model
- If you want to train in Multi Nodes please refer to ./scripts/multi_node_train_*.sh
```
bash 3.single_node_train_gemma.sh
```
(Optional) Proxy-Tuning: Directly improve model capabilities without fine-tuning
```
  bash src/proxy-tuning/scripts/eval/proxy_tuning.sh
```
Evaluate your model: Generate score for benchmark
```
bash 4.eval.sh
```

Evaluate your model: Play with your ckpts in bash

python ./src/evaluate/cli_demo.py --model_name='./ckpts/your/path/tfmr'

Acknowledgment

Citation

Please use the following citation if you intend to use our dataset for training or evaluation:

@misc{wang2024apollo,
   title={Apollo: Lightweight Multilingual Medical LLMs towards Democratizing Medical AI to 6B People},
   author={Xidong Wang and Nuo Chen and Junyin Chen and Yan Hu and Yidong Wang and Xiangbo Wu and Anningzhe Gao and Xiang Wan and Haizhou Li and Benyou Wang},
   year={2024},
   eprint={2403.03640},
   archivePrefix={arXiv},
   primaryClass={cs.CL}
}

FreedomIntelligence/Apollo