Our paper was accepted for the findings of ACL 2024!
Also, we release the enhanced versions of Asclepius: Asclepius-Llama3-8B and Asclepius-Mistral-7B-v0.3.
The pretrained-only versions of Asclepius and Asclepius-R (before instruction tuning) were uploaded to Physionet and Huggingface!
Asclepius-R and the MIMIC-III note instruction are now available on Physionet!
We are excited to introduce the new Asclepius, based on Llama2! You can find these models at the links below:
We introduce the first shareable clinical large language model, Asclepius. The development of large language models tailored for handling patients' clinical notes is often hindered by the limited accessibility and usability of these notes due to strict privacy regulations. To address these challenges, we trained our model, Asclepius, on synthetic clinical notes generated from publicly available case reports extracted from biomedical literature. On GPT-4 and expert evaluation, our model shows comparable performance to the model trained on real clinical notes. The model checkpoints and data are publicly available via huggingface.
prompt = """You are an intelligent clinical languge model.
Below is a snippet of patient's discharge summary and a following instruction from healthcare professional.
Write a response that appropriately completes the instruction.
The response should provide the accurate answer to the instruction, while being concise.
[Discharge Summary Begin]
{note}
[Discharge Summary End]
[Instruction Begin]
{question}
[Instruction End]
"""
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("starmpcc/Asclepius-7B", use_fast=False)
model = AutoModelForCausalLM.from_pretrained("starmpcc/Asclepius-7B")
note = "This is a sample note"
question = "What is the diagnosis?"
model_input = prompt.format(note=note, question=question)
input_ids = tokenizer(model_input, return_tensors="pt").input_ids
output = model.generate(input_ids)
print(tokenizer.decode(output[0]))
Our synthetic discharge summaries, and corresponding instruction-answer pairs are available on huggingface. Due to the license issue of MIMIC-III dataset, instruction-answer pairs used for Asclepius-R is available on Physionet.
Our models, Asclepius-7B and Asclepius-13B is available on huggingface (7B, 13B). Asclepius-R-7B and Asclepius-R-13B are available on Physionet
Environment Setup
conda create -n asclepius python=3.9 -y
conda activate asclepius
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia -y
pip install pandarallel pandas jupyter numpy datasets sentencepiece openai wandb accelerate tiktoken transformers==4.32.0
Note Synthesis
-
Convert PMC-Patients to GPT input
python preprocessing/pmc_preprocessing.py --input_path {PMC-Patients.csv path} --save_path {A}
-
Run OpenAI API
python preprocessing/openai_parallel_request.py --input_path {A} --save_path {B} --request_url {OPENAI_API_ENDPOINT} --api_key {OPENAI_API_KEY}
-
Postprocess OpenAI API output
python preprocessing/pmc_postprocessing.py
--input_path {B} --save_path {C}`
Instruction Generation
-
To use MIMIC-III instead of synthetic notes
python preprocessing/mimiciii_preproc.py --input_path {NOTEEVENTS.csv path} --save_path {C}
-
Make Instructions
python preprocessing/bootstrapping.py --input_path {C} --save_path {D} --request_url {OPENAI_API_ENDPOINT} --api_key {OPENAI_API_KEY}
-
Postprocess Instructions
python preprocesisng/postproc_questions.py --input_path {D} --save_path {E}
-
Make Answers
python preprocessing/openai_parallel_request.py --input_path {E} --save_path {F} --request_url {OPENAI_API_ENDPOINT} --api_key {OPENAI_API_KEY}
-
Postprocess Answers
python preprocessing/make_instruction_dataset.py --input_path {F} --save_path {G}
Pretraining
- Note Preprocessing
- We concatenate all notes with
<eos>
tokens. $ python src/tokenize_data.py --input_path {C} --save_path {H}
- We concatenate all notes with
- Run Pretriaining (for 8x A6000 gpus)
- Adjust
nproc_per_node
,gradient accumulate step
,per_device_train_batch_size
to fit to your hardware (global batch size=128).
$ torchrun --nproc_per_node=8 --master_port={YOUR_PORT} \ src/train.py \ --model_name_or_path "decapoda-research/llama-7b-hf" \ --data_path {H} \ --bf16 True \ --output_dir {I} \ --num_train_epochs 1 \ --per_device_train_batch_size 2 \ --per_device_eval_batch_size 2 \ --gradient_accumulation_steps 8 \ --evaluation_strategy "no" \ --save_strategy "steps" \ --save_steps 1000 \ --learning_rate 2e-5 \ --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 True \ --model_max_length 2048 \ --gradient_checkpointing True
- Adjust
Instruction Finetuning
$ torchrun --nproc_per_node=8 --master_port={YOUR_PORT} \
src/instruction_ft.py \
--model_name_or_path {I} \
--data_path {G} \
--bf16 True \
--output_dir ./checkpoints \
--num_train_epochs 3 \
--per_device_train_batch_size 2 \
--per_device_eval_batch_size 2 \
--gradient_accumulation_steps 8 \
--evaluation_strategy "no" \
--save_strategy "epoch" \
--learning_rate 2e-5 \
--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 True \
--model_max_length 2048 \
--gradient_checkpointing True
--ddp_timeout 18000
Evaluation
-
Evaluate Models
python src/evaluate.py \ --model_name {I} \ --input_path {J} \ --save_path {K}
-
Run GPT-4 for evaluation
python eval/gpt4_evaluate.py --input_path {K} --save_path {L}
@inproceedings{kweon-etal-2024-publicly,
title = "Publicly Shareable Clinical Large Language Model Built on Synthetic Clinical Notes",
author = "Kweon, Sunjun and
Kim, Junu and
Kim, Jiyoun and
Im, Sujeong and
Cho, Eunbyeol and
Bae, Seongsu and
Oh, Jungwoo and
Lee, Gyubok and
Moon, Jong Hak and
You, Seng Chan and
Baek, Seungjin and
Han, Chang Hoon and
Jung, Yoon Bin and
Jo, Yohan and
Choi, Edward",
editor = "Ku, Lun-Wei and
Martins, Andre and
Srikumar, Vivek",
booktitle = "Findings of the Association for Computational Linguistics ACL 2024",
month = aug,
year = "2024",
address = "Bangkok, Thailand and virtual meeting",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.findings-acl.305",
pages = "5148--5168",
abstract = "The development of large language models tailored for handling patients{'} clinical notes is often hindered by the limited accessibility and usability of these notes due to strict privacy regulations.To address these challenges, we first create synthetic large-scale clinical notes using publicly available case reports extracted from biomedical literature.We then use these synthetic notes to train our specialized clinical large language model, Asclepius.While Asclepius is trained on synthetic data, we assess its potential performance in real-world applications by evaluating it using real clinical notes.We benchmark Asclepius against several other large language models, including GPT-3.5-turbo and other open-source alternatives. To further validate our approach using synthetic notes, we also compare Asclepius with its variants trained on real clinical notes. Our findings convincingly demonstrate that synthetic clinical notes can serve as viable substitutes for real ones when constructing high-performing clinical language models. This conclusion is supported by detailed evaluations conducted by both GPT-4 and medical professionals. All resources{---}including weights, codes, and data{---}used in the development of Asclepius will be made publicly accessible for future research.",
}