pixas/MedSSS_Policy

MedSSS-8B-Policy

GitHub | Paper

Introduction

MedSSS-Policy is a the policy model designed for slow-thinking medical reasoning. It will conduct explicit step-wise reasoning and finalize the answer at the end of the response.

For more information, visit our GitHub repository: https://github.com/pixas/MedSSS.

Usage

We build the policy model as a LoRA adapter, which saves the memory to use it. As this LoRA adapter is built on Meta-Llama3.1-8B-Instruct, you need to first prepare the base model in your platform. You can deploy it with tools like vllm or Sglang, or perform direct inference:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct",torch_dtype="auto",device_map="auto")
model = PeftModel.from_pretrained(base_model, "pixas/MedSSS_Policy", torc_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("pixas/MedSSS_Policy")
input_text = "How to stop a cough?"
messages = [{"role": "user", "content": input_text}]
inputs = tokenizer(tokenizer.apply_chat_template(messages, tokenize=False,add_generation_prompt=True
), return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

MedSSS-Policy adopts a step-wise reasoning approach, with outputs formatted as:

Step 0: Let's break down this problem step by step.
Step 1: ...
[several steps]
Step N: [last reasoning step]\n\nThe answer is {answer}