Democratizing Medical LLMs For Much More Languages

Covering 12 Major Languages including English, Chinese, French, Hindi, Spanish, Arabic, Russian, Japanese, Korean, German, Italian, Portuguese and 38 Minor Languages So far.

📃 Paper • 🌐 Demo • 🤗 ApolloMoEDataset • 🤗 ApolloMoEBench • 🤗 Models • 🌐 Apollo

🌈 Update

• [2024.10.15] ApolloMoE repo is published！🎉

Languages Coverage

12 Major Languages and 38 Minor Languages

Click to view the Languages Coverage

Architecture

Click to view the MoE routing image

Results

Dense

🤗 Apollo2-0.5B • 🤗 Apollo2-1.5B • 🤗 Apollo2-2B • 🤗 Apollo2-3.8B • 🤗 Apollo2-7B • 🤗 Apollo2-9B

Click to view the Dense Models Results

Post-MoE

🤗 Apollo-MoE-0.5B • 🤗 Apollo-MoE-1.5B • 🤗 Apollo-MoE-7B

Click to view the Post-MoE Models Results

Usage Format

Apollo2

• 0.5B, 1.5B, 7B: User:{query}\nAssistant:{response}<|endoftext|>
• 2B, 9B: User:{query}\nAssistant:{response}<eos>
• 3.8B: <|user|>\n{query}<|end|><|assisitant|>\n{response}<|end|>

Apollo-MoE

• 0.5B, 1.5B, 7B: User:{query}\nAssistant:{response}<|endoftext|>

Dataset & Evaluation

• Dataset 🤗 ApolloMoEDataset

  Click to expand

  

  • Data category

  The complete data is stored in ApolloMoEDataset.json, while a sample shown in ApolloMoEDataset_sample.json

• Evaluation 🤗 ApolloMoEBench

  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
    • [MMLU_FR]
      • Clinical knowledge, Medical genetics, Anatomy, Professional medicine, College biology, College medicine

  • HI: MMLU_HI

    • Clinical knowledge, Medical genetics, Anatomy, Professional medicine, College biology, College medicine

  • AR: MMLU_AR

    • Clinical knowledge, Medical genetics, Anatomy, Professional medicine, College biology, College medicine

  • JA: IgakuQA

  • KO: KorMedMCQA

  • IT:

    • MedExpQA
    • [MMLU_IT]
      • Clinical knowledge, Medical genetics, Anatomy, Professional medicine, College biology, College medicine

  • DE: BioInstructQA: German part

  • PT: BioInstructQA: Portuguese part

  • RU: RuMedBench

Model Download and Inference

We take Apollo-MoE-0.5B as an example

1. Login Huggingface

   huggingface-cli login --token $HUGGINGFACE_TOKEN
   

2. Download model to local dir

   from huggingface_hub import snapshot_download
   import os
   
   local_model_dir=os.path.join('/path/to/models/dir','Apollo-MoE-0.5B')
   snapshot_download(repo_id="FreedomIntelligence/Apollo-MoE-0.5B", local_dir=local_model_dir)

3. Inference Example

   from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
   import os
   
   local_model_dir=os.path.join('/path/to/models/dir','Apollo-MoE-0.5B')
   
   model=AutoModelForCausalLM.from_pretrained(local_model_dir,trust_remote_code=True)
   tokenizer = AutoTokenizer.from_pretrained(local_model_dir,trust_remote_code=True)
   generation_config = GenerationConfig.from_pretrained(local_model_dir, pad_token_id=tokenizer.pad_token_id, num_return_sequences=1, max_new_tokens=7, min_new_tokens=2, do_sample=False, temperature=1.0, top_k=50, top_p=1.0)
   
   inputs = tokenizer('Answer direclty.\nThe capital of Mongolia is Ulaanbaatar.\nThe capital of Iceland is Reykjavik.\nThe capital of Australia is', return_tensors='pt')
   inputs = inputs.to(model.device)
   pred = model.generate(**inputs,generation_config=generation_config)
   print(tokenizer.decode(pred.cpu()[0], skip_special_tokens=True))

Results reproduction

(Optional) Custom Model as Base

Click to expand

   copy /path/to/your/configuration_upcycling_qwen2_moe.py /path/to/src/variants/moe_initilization/configuration_upcycling_qwen2_moe.py
   copy /path/to/your/modeling_upcycling_qwen2_moe.py /path/to/src/variants/moe_initilization/modeling_upcycling_qwen2_moe.py
   cd /path/to/src/variants/moe_initilization
   bash convert.sh

Full-finetune on Base Model

Click to expand

We take Apollo2-7B or Apollo-MoE-0.5B as examples

1. Download and extract data:

   • Dowload Dataset and Benchmark firstly
   • Extract major or minor data part according to your needs:

   bash 0.extract_data.sh
   

2. Prepare test and dev data for specific model:

   • Create test data for with special token

   bash 1.data_process_test&dev.sh
   

3. 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
   

4. Train the model

   • If you want to train in Multi Nodes please refer to ./src/sft/training_config/zero_multi.yaml

   bash 3.single_node_train.sh
   

5. Evaluate your model: Generate score for benchmark

   bash 4.eval.sh
   

Citation

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

@misc{zheng2024efficientlydemocratizingmedicalllms,
      title={Efficiently Democratizing Medical LLMs for 50 Languages via a Mixture of Language Family Experts}, 
      author={Guorui Zheng and Xidong Wang and Juhao Liang and Nuo Chen and Yuping Zheng and Benyou Wang},
      year={2024},
      eprint={2410.10626},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.10626}, 
}