Yunxiang Li1, Zihan Li2, Kai Zhang3, Ruilong Dan4, You Zhang1
100k real conversations between patients and doctors from HealthCareMagic.com HealthCareMagic-100k.
15k real conversations between patients and doctors from icliniq.com icliniq-15k.
Checkpoints of ChatDoctor, fill this form.
Stanford Alpaca data for basic conversational capabilities. Alpaca link.
In a conda env with pytorch available, run:
pip install -r requirements.txt
Demo Page: https://huggingface.co/spaces/ChatDoctor/ChatDoctor It is worth noting that our model has not yet achieved 100% accurate output, please do not apply it to real clinical scenarios.
For those who want to try the online demo, please register for hugging face and fill out this form link.
You can download the following training dataset
100k real conversations between patients and doctors from HealthCareMagic.com HealthCareMagic-100k.
15k real conversations between patients and doctors from icliniq.com icliniq-15k.
5k generated conversations between patients and physicians from ChatGPT GenMedGPT-5k and disease database.
Our model was firstly be fine-tuned by Stanford Alpaca's data to have some basic conversational capabilities. Alpaca link
In order to download the checkpoints, fill this form: link. Place the model weights file in the ./pretrained folder.
torchrun --nproc_per_node=4 --master_port=<your_random_port> train.py \
--model_name_or_path <your_path_to_hf_converted_llama_ckpt_and_tokenizer> \
--data_path ./HealthCareMagic-100k.json \
--bf16 True \
--output_dir pretrained \
--num_train_epochs 1 \
--per_device_train_batch_size 4 \
--per_device_eval_batch_size 4 \
--gradient_accumulation_steps 8 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 2000 \
--save_total_limit 1 \
--learning_rate 2e-6 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--fsdp "full_shard auto_wrap" \
--fsdp_transformer_layer_cls_to_wrap 'LLaMADecoderLayer' \
--tf32 TrueFine-tuning with Lora
WORLD_SIZE=6 CUDA_VISIBLE_DEVICES=0,1,2,3,4,5 torchrun --nproc_per_node=6 --master_port=4567 train_lora.py \
--base_model './weights-alpaca/' \
--data_path 'HealthCareMagic-100k.json' \
--output_dir './lora_models/' \
--batch_size 32 \
--micro_batch_size 4 \
--num_epochs 1 \
--learning_rate 3e-5 \
--cutoff_len 256 \
--val_set_size 120 \
--adapter_name loraYou can build a ChatDoctor model on your own machine and communicate with it.
python chat.pyWe fine-tuned our model on HealthCareMagic-100K and GenMedGPT-5k, and test our model on icliniq-15k.
Patient question from icliniq-15k: Hi doctor,I am getting one-sided headache. I do not get them often. They are behind my right eye and temple area. Now this time I am also fighting with heavy cold and cough. It came on late morning today and it hurts more when I cough or shake my head or bend over. Prior to this, I had the same headache about four months ago. At that time it came on midafternoon and lasted till next morning. Prior to that, I had the same headache exactly on the right side and lasted for a day. I am on CCB and Protonix. I am scared of brain tumor. Please help
Real doctor answer from icliniq-15k: Hi. Based on what you have told, this is more likely to be due to sinusitis. Sinuses are small air filled cavities in your skull bone. They can get infected and when they are infected, you will get headaches behind eyes. Such headaches will be more common during cold and cough. Brain tumor headache will be on both sides, more in forehead and you will also have vomiting in that case. Based on what you have told, this is more likely due to simple sinus infection instead of brain tumor. CT brain and CT paranasal sinuses.
Answer form ChatGPT: Hello! I'm sorry to hear you're experiencing these headaches. Based on the location of the pain, it sounds like it may be a tension headache or a migraine. These types of headaches can sometimes be triggered by colds or other illnesses. However, it's always best to get checked out by your doctor to rule out any other underlying conditions. In terms of a brain tumor, that is a very rare cause of headaches. Most people who experience headaches do not have a brain tumor. That being said, if you have any concerning symptoms such as vision changes, weakness, or numbness, it is important to seek medical attention right away. In the meantime, you can try taking over-the-counter pain relievers such as ibuprofen or acetaminophen to help with the headache. It's also important to get plenty of rest and stay hydrated. If the headaches persist or become more frequent, please make an appointment with your doctor. Take care!
Answer form ChatDoctor: Hi, Thank you for posting your query. First, I would like to reassure you that there is no need to worry about brain tumor. Also, we should keep migraine apart as well. The most likely cause of your headaches could be sinusitis. This can be confirmed by doing an X-ray of face. Treatment would consist of using decongestant tablets/ nasal sprays. Best wishes, Chat Doctor.
ChatDoctor accurately mentions sinusitis, which is also present in the real conversation in icliniq-15k.
ChatDoctor is a next-generation AI doctor model that is based on the LLaMA model. The goal of this project is to provide patients with an intelligent and reliable healthcare companion that can answer their medical queries and provide them with personalized medical advice.
The ChatDoctor is an advanced language model that is specifically designed for medical applications. It has been trained on a large corpus of medical literature and has a deep understanding of medical terminology, procedures, and diagnoses. This model serves as the foundation for ChatDoctor, enabling it to analyze patients' symptoms and medical history, provide accurate diagnoses, and suggest appropriate treatment options.
The ChatDoctor model is designed to simulate a conversation between a doctor and a patient, using natural language processing (NLP) and machine learning techniques. Patients can interact with the ChatDoctor model through a chat interface, asking questions about their health, symptoms, or medical conditions. The model will then analyze the input and provide a response that is tailored to the patient's unique situation.
One of the key features of the ChatDoctor model is its ability to learn and adapt over time. As more patients interact with the model, it will continue to refine its responses and improve its accuracy. This means that patients can expect to receive increasingly personalized and accurate medical advice over time.
Recent large language models (LLMs) in the general domain, such as ChatGPT, have shown remarkable success in following instructions and producing human-like responses. However, such language models have not been tailored to the medical domain, resulting in poor answer accuracy and inability to give plausible recommendations for medical diagnosis, medications, etc. To address this issue, we collected more than 700 diseases and their corresponding symptoms, required medical tests, and recommended medications, from which we generated 5K doctor-patient conversations. In addition, we obtained 100K real patient-doctor conversations from online Q&A medical consultation sites. By fine-tuning LLMs using these doctor-patient conversations, the resulting models emerge with great potential to understand patients' needs, provide informed advice, and offer valuable assistance in a variety of medical-related fields. The integration of these advanced language models into healthcare can revolutionize the way healthcare professionals and patients communicate, ultimately improving the overall efficiency and quality of patient care and outcomes. In addition, we made public all the source codes, datasets, and model weights to facilitate the further development of dialogue models in the medical field.
The development of instruction-following large language models (LLMs) such as ChatGPT has garnered significant attention due to their remarkable success in instruction understanding and human-like response generation. These auto-regressive LLMs are pre-trained over web-scale natural languages by predicting the next token and then fine-tuned to follow large-scale human instructions. Also, they have shown strong performances over a wide range of NLP tasks and generalizations to unseen tasks, demonstrating their potential as a unified solution for various problems such as natural language understanding, text generation, and conversational AI. However, the exploration of such general-domain LLMs in the medical field remains relatively untapped, despite the immense potential they hold for transforming healthcare communication and decision-making. The specific reason is that the existing models do not learn the medical field in detail, resulting in the models often giving wrong diagnoses and wrong medical advice when playing the role of a doctor. By fine-tuning the large language dialogue model on the data of doctor-patient conversations, the application of the model in the medical field can be significantly improved. Especially in areas where medical resources are scarce, ChatDoctor can be used for initial diagnosis and triage of patients, significantly improving the operational efficiency of existing hospitals.
Since large language models such as ChatGPT are in a non-open source state, we used Meta's LLaMA and first trained a generic conversation model using 52K instruction-following data provided by Stanford Alpaca, and then fine-tuned the model on our collected physician-patient conversation dataset. The main contributions of our method are three-fold:
- We designed a process framework for fine-tuning large language models in the medical domain.
- We collected a dataset with 5,000 generated doctor-patient conversations and 200,000 real patient-doctor conversations for fine-tuning the large language model.
- We validate that the fine-tuned bigrams with medical domain knowledge have real potential for clinical application.
The first step in building a physician-patient conversation dataset is to collect the disease database that serves as the gold standard. Therefore, we collected and organized a database of diseases, which contains about 700 diseases with their relative symptoms, medical tests, and recommended medications. To train high-quality conversation models on an academic budget, we input each message from the disease database separately as a prompt into the ChatGPT API to automatically generate instruction data. It is worth noting that our prompts to the ChatGPT API contain the gold standard of diseases and symptoms, and drugs, so our fine-tuned ChatDoctor is not only able to achieve ChatGPT's conversational fluency but also higher diagnostic accuracy compared to ChatGPT. We finally collected 5K doctor-patient conversation instructions and named it InstructorDoctor-5K.
The generated conversations, while ensuring accuracy, have a low diversity of conversations. Therefore, we also collected about 100k real doctor-patient conversations from an online Q&A based medical advisory service website -- "Health Care Magic." We manually and automatically filtered these data to remove physician and patient names and used language tools to correct grammatical errors in the responses.
We build ChatDoctor utilizing Meta's LLaMA model, a distinguished publicly accessible LLM. Notably, in spite of its 7 billion parameters, LLaMA has been reported that LLaMA's efficacy can attain competitive or superior outcomes in comparison to the considerably larger GPT-3 (with 175 billion parameters) on several NLP benchmarks. LLaMA's performance improvement was achieved by amplifying the magnitude of training data, as opposed to parameter quantity. Specifically, LLaMA was trained on 1.4 trillion tokens, procured from publicly accessible data repositories such as CommonCrawl and arXiv documents. We utilize conversation demonstrations synthesized via ChatGPT and subsequently validated by medical practitioners to fine-tune the LLaMA model, in accordance with the Stanford Alpaca training methodology, and our model was firstly be fine-tuned by Stanford Alpaca's data to have some basic conversational capabilities. The fine-tuning process was conducted using 6 A*100 GPUs for a duration of 30 minutes. The hyperparameters employed in the training process were as follows: the total batch size of 192, a learning rate of 2e-5, a total of 3 epochs, a maximum sequence length of 512 tokens, a warmup ratio of 0.03, with no weight decay.
We emphasize that ChatDoctor is for academic research only and any commercial use and clinical use is prohibited. There are three factors in this decision: First, ChatDoctor is based on LLaMA and has a non-commercial license, so we necessarily inherited this decision. Second, our model is not licensed for healthcare-related purposes. Also, we have not designed sufficient security measures, and the current model still does not guarantee the full correctness of medical diagnoses.
ChatDoctor: A Medical Chat Model Fine-tuned on LLaMA Model using Medical Domain Knowledge
@misc{yunxiang2023chatdoctor,
title={ChatDoctor: A Medical Chat Model Fine-tuned on LLaMA Model using Medical Domain Knowledge},
author={Li, Yunxiang and Li, Zihan and Zhang, Kai and Dan, Ruilong and Zhang, You},
year={2023},
eprint={2303.14070},
archivePrefix={arXiv},
primaryClass={cs.CL}
}