Apollo: Lightweight Multilingual Medical LLMs

https://arxiv.org/abs/2311.09774

 

HuatuoGPT-II, One-stage Training for Medical Adaption of LLMs

 

GOAL : Aims to develop medical LLMs across the six most widely spoken languages.

Launched in a variety of relatively small sizes (e.g., 0.5B, 1.8B, 2B, 6B, 7B) on multilingual healthcare benchmarks, Apollo models achieve the best performance among same size models.

pip install bitsandbytes

pip install accelerate

 

import torch

from transformers import AutoModelForCausalLM, AutoTokenizer

from transformers.generation.utils import GenerationConfig

tokenizer = AutoTokenizer.from_pretrained("FreedomIntelligence/HuatuoGPT-2-7b", use_fast=True, trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained("FreedomIntelligence/HuatuoGPT-2-7b", device_map="auto", torch_dtype=torch.bfloat16, trust_remote_code=True)

model.generation_config = GenerationConfig.from_pretrained("FreedomIntelligence/HuatuoGPT-2-7b")

messages = []

messages.append({"role": "user", "content": "肚子疼怎么办？"})

response = model.HuatuoChat(tokenizer, messages)

print(response)

 

Currently, my laptop is equipped with torch, TensorFlow, and AI environment, but there is no free space, so I decided to use colab.

 used colab to implement quick code but I couldn't use it with a free version.

https://colab.research.google.com/drive/1Q-Dnrs8pqIgwfn4YFqH6ETzvy0N14e_H

Google Colaboratory Notebook

 

So I found a proxy server that could use that model and tried it.

https://www.huatuogpt.cn/#/

 

Review of use:

weakness : The questions were well recognized in six languages. However, they answered only in Chinese.

strength : It can improve the multilingual healthcare capabilities of larger models without fine-tuning.

opportunity : Increased access to healthcare worldwide

Threats: If the model has a technical fault, the patient will have a problem