Gemma-2-9b-medical

Introduction

This repo contains Gemma-2-9b-Medical, a medical language model with 9 billion parameters. This model builds upon the foundation of Gemma-2-9b-base and has been tuned with diverse medical and general instructions. We also use the three strategies in the paper 'Efficient Continual Pre-training by Mitigating the Stability Gap' to mitigate the stability gap during instruction tuning, which boosts the model's medical task performance and reduces the computation consumption.

💻 Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch
model_name = "YiDuo1999/Gemma-2-9b-medical"
device_map = 'auto'
model = AutoModelForCausalLM.from_pretrained( model_name, trust_remote_code=True,use_cache=False,device_map=device_map)
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token
def askme(question):
    sys_message = ''' 
    You are an AI Medical Assistant trained on a vast dataset of health information. Please be thorough and
    provide an informative answer. If you don't know the answer to a specific medical inquiry, advise seeking professional help.
    '''   
    # Create messages structured for the chat template
    messages = [{"role": "system", "content": sys_message}, {"role": "user", "content": question}]
    
    # Applying chat template
    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, max_new_tokens=100, use_cache=True)
    
    # Extract and return the generated text, removing the prompt
    response_text = tokenizer.batch_decode(outputs)[0].strip()
    answer = response_text.split('<|im_start|>assistant')[-1].strip()
    return answer

🏆 Evaluation

For question-answering tasks, we have

Citation

@inproceedings{Guo2024EfficientCP,
  title={Efficient Continual Pre-training by Mitigating the Stability Gap},
  author={Yiduo Guo and Jie Fu and Huishuai Zhang and Dongyan Zhao and Yikang Shen},
  year={2024},
  url={https://api.semanticscholar.org/CorpusID:270688100}
}