简介（Introduction）

海智若康大模型，面向日常健康咨询领域，提供7*24小时实时问答咨询，助力实现“主动健康”。该模型在Qwen2.5的基础上，整合海量医学知识和医疗数据训练而来，在语言理解、医学知识等多方面能力均有大幅提升。 本次开源内容含模型权重文件和推理代码。

HalerockHealth Large Language Model is aimed at daily health consultation, providing 24/7 real-time Q&A support to help achieve the goal of "health initiative". Built upon Qwen2.5, the model integrates extensive medical knowledge and leverages large-scale medical data for training, significantly enhancing its capabilities in language comprehension, medical expertise, and other areas. This open-source content includes model weights and inference code.

(Built with Qwen)

模型效果评测（Evaluation）

测试数据（Testing Data）

该模型在多种基准测试上进行了评估，涵盖模型通用能力、医学基础知识和医患咨询能力。主要的评估基准包括：

1. 模型通用能力：包括arc-c、arc-e、bbh、c3和ifeval，旨在评估模型的推理能力、语言理解能力和指令遵循能力；
2. 医学知识：包括CMExam和CMB，旨在评估模型的医学知识储备；
3. 医患咨询能力：包括CMtMedQA、WebMedQA和HuaTuo共三个数据集的测试集，要求模型补全对话，并使用大模型从安全性、专业性和流畅性三个角度评估模型输出与参考回答的优胜率。

The model has been evaluated on multiple benchmarks, covering general capabilities, medical knowledge, and doctor-patient consultation skills. The main evaluation benchmarks include:

1. General Capabilities: Including arc-c, arc-e, bbh, c3, and ifeval, which aim to assess the model's reasoning ability, language comprehension, and instruction-following ability.
2. Medical Knowledge: Including CMExam and CMB, which aim to evaluate the model's medical knowledge.
3. Doctor-Patient Consultation Skills: Including the test sets of three datasets, CMtMedQA, WebMedQA, and HuaTuo, which require the model to complete dialogues and assess the model's output against reference answers from the perspectives of safety, professionalism, and fluency.

效果（Preference）

	Citrus1.0-Qwen-72B	Qwen2.5-72B	Qwen2.5-72B-Instruct	HalerockHealth
arc-c	-	85.67	95.48	95.90
arc-e	-	92.90	98.70	98.57
bbh	51.39	83.97	87.01	85.78
c3	14.25	96.00	95.70	96.26
ifeval(loose)	66.54	57.12	85.21	85.21
ifeval(strict)	57.86	50.28	83.36	80.78
CMExam	87.32	79.93	83.95	89.90
CMB	83.60	76.28	82.36	87.04
CMtMedQA	90.48	29.58	88.89	93.77
WebMedQA	99.50	89.25	98.50	100
HuaTuo	99	88.50	97.50	98.75

用例（Usage）

可以通过 vllm 等工具来部署，或者直接进行推理:

You can deploy it with tools like vllm, or perform direct inference:

from transformers import AutoModelForCausalLM, AutoTokenizer
model_path = "The full path of the model"

model = AutoModelForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)

input_text = "怎样预防感冒？"
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))

许可证（License）

1.《开放原子模型许可证 第一版》 《OpenAtom Model License, Version 1.0》

2.Qwen LICENSE AGREEMENT

免责声明（Disclaimer）

海智若康大模型旨在为用户提供健康相关信息。然而，该大模型生成的信息并非旨在替代专业的医疗建议、诊断或治疗。因此，如有任何医疗问题或疑问，请务必咨询持牌医疗保健提供者。本项目对因使用该模型而造成的任何伤害、损失或损害不承担任何责任。

HalerockHealth Large Language Model is a tool designed to provide health-related information to users. However, the information generated by the model is not intended to be a substitute for professional medical advice, diagnosis, or treatment. Therefore, it is essential to consult a licensed healthcare provider for any medical concerns or questions. This project disclaims all liability for any injuries, losses, or damages arising from the use of this model.